Quantitative Methodologies using Multi-Methods: Models for Social Science and Information Technology Research 9780367903961, 9781032046976, 9781003024149


309 127 9MB

English Pages 292 [311] Year 2021

Report DMCA / Copyright

DOWNLOAD PDF FILE

Table of contents :
Cover
Half Title
Title Page
Copyright Page
Contents
Preface: Possible Uses of this Book
Introduction
SECTION I: Development of the Methodological Modules
Chapter 1: Pre-Requisite General Questions
Impact of the Assumption of Homogeneity of the Sample on Research Questions
From a Basket of Apples to a Set of Systems (Decision-Making Units)
From Systems to Systems in Context
Chapter 2: Components of Multi-Method Methodologies
Cluster Analysis (CA)
Classification Decision Trees Induction (CDTI)
Neural Networks (NNs)
Association Rules Mining (ARM)
Data Envelopment Analysis (DEA)
Multiple Regression (MR)
Chapter 3: Framework for Methodological Modules
SECTION II: Description of the Methodological Modules
Chapter 4: A1: Homogeneous Sample – DEA and DTI
Phase 1: DEA
Phase 2: DTI
Examples of Application of DEA and DTI
Chapter 5: A2: Homogeneous Sample – DEA and ARM
Phase 1: DEA
Phase 2: ARM
Examples of Application of DEA and ARM
Chapter 6: B1: Heterogeneous Sample (Groupings Are Given) – DTI and ARM
Phase 1: DTI
Phase 2: ARM
Examples of Application of DTI and ARM
Chapter 7: B2: Heterogeneous Sample (Groupings Are Given) – DTI and MR
Phase 1: DTI
Option 1: DTI Using the Data Set Comprised of a Causal Model Only
Option 2: DTI Using the Data Set without Causal Model
Option 3: DTI Using the Complete Data Set
Phase 2: MR
Option 1: MR Using the Causal Model Only
Option 2: MR Using the Adapted Causal Model – Contextual Independent Variable
Option 3: Creating a New MR Using Contextual Independent Variables
Example of Application of DTI and MR
Chapter 8: B3: Heterogeneous Sample (Groupings Are Given) – DTI, DEA, and ARM
Phase 1: DTI
Option 1: The Data Set Is Comprised of the Variables of the DEAea MODel
Option 2: The Data Set Contains Contextual Variables
Phase 2: DEA
Phase 3: ARM
Option 1: ARM to Generate “If→ (Level of the Top-Split Variable(s))”
Option 2: ARM to Generate “If→ (DEA Model’s Inputs)”
Option 3: ARM to Generate “If→ (DEA Model’ s Outputs)”
Option 4: ARM to Generate “If→ (Level of Averaged Relative Efficiency)”
Option 5: ARM to Generate “If→ (Received Categorization)”
Examples of Application of DTI, DEA, and ARM
Chapter 9: B4: Heterogeneous Sample (Groupings Are Given) – DTI, DEA, and NN
Phase 1: DTI
Phase 2: DEA
Phase 3: NN
Step 1: Generate NN Model of Transformative Capacity
Step 2: Generate Outputs of a Less Efficient Group Based on Transformative Capacity of a More Efficient Group
Step 3: Generate Outputs of a More Efficient Group Based on Transformative Capacity of a Less Efficient Group
Step 4: Compile the Generated Outputs in a New Data Set
Phase 4: DEA
Example of Application of DTI, DEA, and NN
Chapter 10: C1: Heterogeneous Sample (Groupings Are Not Known) – CA and DTI
Phase 1: CA
Phase 2: DTI
Examples of Application of CA and DTI
Chapter 11: C2: Heterogeneous Sample (Groupings Are Not Known) – CA and ARM
Phase 1: CA
Phase 2: ARM
Option 1: ARM Using Only Intrinsic Variables
Option 2: ARM Using Only Contextual Variables
Option 3: ARM Using Intrinsic and Contextual Variables
Examples of Application of CA and ARM
Chapter 12: C3: Heterogeneous Sample (Groupings Are Not Known) – CA, DTI, and MR
Phase 1: CA
Phase 2: DTI
Option 1: Data Set Is Limited to Variables of the MR Model
Option 2: Data Set Comprises Variables of the MR Model and Contextual Variables
Phase 3: MR
Example of Application of CA, DTI, and MR
Chapter 13: C4: Heterogeneous Sample (Groupings Are Not Known) – CA, DTI, and ARM
Phase 1: CA
Phase 2: DTI
Option 1: A Priori Target Variable
Option 2: CA-based Target Variable
Phase 3: ARM
Step 1
Step 2
Step 3
Step 4
Examples of Application of CA, DTI, and ARM
Chapter 14: C5: Heterogeneous Sample (Groupings Are Not Known) – CA and DEA
Phase 1: CA
Option 1: CA based on the DEA Model
Option 2: CA based on the DEA Model and Contextual Variables
Phase 2: DEA
Examples of Application of CA and DEA
Chapter 15: C6: Heterogeneous Sample (Groupings Are Not Known) – CA, DEA, and ARM
Phase 1: CA
Phase 2: DEA
Phase 3: ARM
Option 1: Complete Sample, # of Variables = the DEA Model
Option 2: Complete Sample, # of Variables = the DEA Model + Contextual Variables
Option 3: Sub-Sets of the Sample, # of Variables = the DEA Model
Option 4: Sub-sets of the Sample, # of Variables = the DEA Model + Contextual Variables
Examples of Application of CA, DEA, and ARM
Chapter 16: C7: Heterogeneous Sample (Groupings Are Not Known) – CA, DTI, and DEA
Phase 1: CA
Phase 2: DTI
Phase 3: DEA
Examples of Application of CA, DTI, and DEA
Chapter 17: C8: Heterogeneous Sample (Groupings Not Known) – CA, DTI, DEA, and NN
Phase 1: CA
Phase 2: DTI
Phase 3: DEA
Phase 4: NN
Step 1: Creating an NN Model of “Low-Level” Cluster
Step 2: Creating an NN Model of “High-Level” Cluster
Step 3: Simulation of the Outputs of “Low-Level” Cluster Using NN Model of “High-Level” Cluster
Step 4: Simulation of the Outputs of “High-Level” Cluster Using NN Model of “Low-Level” Cluster
Phase 5: DEA
Examples of Application of CA, DTI, DEA, and NN
SECTION III: Methodological Modules – Examples of Their Application
Chapter 18: A Hybrid DEA/DM-based DSS for Productivity-Driven Environments
Introduction
Description of the DSS
Externally Oriented Functionality
Internally Oriented Functionality
Architecture of the DSS
An Illustrative Application
Step 1: Is the Business Environment Homogeneous?
Step 2: What Are the Factors Responsible for Heterogeneity of the Business Environment?
Step 3: Do Groups of Competitors Differ in Terms of the Relative Efficiency?
Step 4: What Are some of the Factors Associated with the Differences in Relative Efficiency?
Step 5: Are There any Complementarities Between the Relevant Variables?
Step 6: What Is a Better Way to Improve Production of Outputs?
Conclusion
Acknowledgment
References
Chapter 19: Determining Sources of Relative Inefficiency in Heterogeneous Samples: Methodology Using Cluster Analysis, DEA, and Neural Networks
Introduction
Description of the Methodology
Description of Steps 3–5 of the Methodology
Step 3: Generate a “Black Box” Model of Transformative Capacity of Each Cluster
Step 4: Generate Simulated Sets of the Outputs for Each Cluster
Step 5: Determine the Sources of the Relative Inefficiency of the DMUs in the Sample
Motivation for Steps 3 and 5 of the Methodology
Motivation for Step 3
Motivation for Step 5
Illustrative Example
Description of the Illustrative Data Set
Application of the Methodology on the Illustrative Data Set
Results of Step 1: Evaluate the Scale Heterogeneity Status of the Data Set
Results of Step 2: Determine the Relative Efficiency Status of Each DMU
Results of Steps 3 and 4: Generate Simulated Sets of the Outputs for Each Cluster Based on Black Box Models Transformative Capacity Processes
Results of Step 5
Discussion and Conclusion
Acknowledgment
References
Chapter 20: Exploring Context Specific Micro-Economic Impacts of ICT Capabilities
Introduction
Theoretical Framework and the Research Model
The Methodology of the Study
Phase 1: Application of Data Envelopment Analysis (DEA)
Phase 1, Step 1
Phase 1, Step 2
Phase 1, Step 3
Phase 2: Decision Tree-Based Analysis
Phase 2, Step 1
Phase 2, Step 2
Description of the Data
Results of the Data Analysis
Results from Phase 1: Application of Data Envelopment Analysis (DEA)
Phase 1, Step 1
Phase 1, Step 2
Phase 1, Step 3
Results from Phase 2 – Decision Tree (DT) Based Analysis
Conclusion
Contributions to Theory
Contributions to Practice
Acknowledgment
References
Chapter 21: A Methodology for Identifying Sources of Disparities in the Socio-Economic Outcomes of ICT Capabilities in SSAs
Introduction
Research Framework
Proposed Methodology
A New Methodology: Benefits and Justifications
Phase 1: Data Envelopment Analysis (DEA)
Phase 2: Decision Tree Induction (DTI)
Phase 3: Association Rule Mining (ARM)
Research Questions and Null Hypotheses of the Study
The Data
Results of the Data Analysis
Phase 1: Data Envelopment Analysis
Phase 2: Decision Tree Induction
Phase 3: Association Rule Mining
Discussion of the Results
Conclusion
Acknowledgment
References
Chapter 22: Discovering Common Causal Structures that Describe Context-Diverse Heterogeneous Groups
Introduction
A Conceptualization of the Benchmarking Problem
Research Problem and Research Questions of the Study
The Proposed Methodology
Description of the Methodology
Justification & Benefits of the Methodology
Illustrative Example – Application to Sub-Saharan Economies
Phase 1: Define the Transformation Framework
Phase 2: Partition the Set of Decision Making Units into Meaningful Groups
Phase 3: Data Envelopment Analysis
Phase 4: Decision Tree Induction (DTI)
Phase 5: Association Rule Mining
Conclusion
Acknowledgment
References
Chapter 23: An Empirical Investigation of ICT Capabilities and the Cost of Business Start-up Procedures in Sub-Saharan African Economies
The Research Framework and Research Questions
Proposed Methodology
Phase 1: Cluster Analysis (CA)
Phase 2: Decision Tree Induction
Phase 3: Data Envelopment Analysis
Phase 4: Ordinary Least Squares Regression
Phase 5: Association Rule Mining
Data
Results of the Data Analysis
Phase 1: CA
Phase 2: DTI
Phase 3: DEA
Phase 4: OLS
Phase 5: ARM
Interpretation of the Results of the Data Analysis
Cluster Analysis
Decision Tree Induction
DEA
Ordinary Least Squares (OLS)
ARM
Discussion of the Results of the Study
Conclusion
Acknowledgment
References
Chapter 24: Exploring the Socio-Economic Impacts of ICT-Enabled Public Value in Sub-Saharan Africa
Introduction
Research Framework of the Study
Research Questions of the Study
Methodology of the Investigation
Phase 1: Cluster Analysis
Phase 2: Decision Tree Induction
Phase 3: Data Envelopment Analysis (DEA)
Phase 4: Ordinary Least Squares (OLS) Regression
Phase 5: Association Rule Mining (ARM)
Data
Results of the Data Analysis
Phase 1: CA
Phase 2: DTI
Phase 3: DEA
Phase 4: OLS
Phase 5: ARM
Discussion of the Results of the Study
Conclusion
Acknowledgment
References
Chapter 25: Contributing Factors to Information Technology Investment Utilization in Transition Economies: An Empirical Investigation
Introduction
Theoretical Framework
Growth Accounting
Theory of Complementarity
Overview on the Data
Methodology: Searching for the Determinants of the Efficiency of Utilization of Investments in Telecoms
Phase 1: Data Envelopment Analysis
Data Used to Perform DEA
Phase 2: Cluster Analysis
Data Used to Perform CA
Phase 3: Decision Tree
Data Used to Perform DT
Results
Results: DEA
Results: Cluster Analysis
Results: Decision Tree
Contribution of the Study
Summary and Conclusion
Acknowledgment
References
Appendix A
Chapter 26: Increasing the Discriminatory Power of DEA in the Presence of the Sample Heterogeneity with Cluster Analysis and Decision Trees
Introduction
The Proposed Methodology
Overview of Data Set of Illustrative Example
Description of the Methodology
Step 1: Determine the Structural Homogeneity Status of the Data Set
Step 2: Determine the Relative Efficiency Status of DMUs
Step 3: Describe the Relative Efficiency Categories
Conclusion
Acknowledgment
References
Chapter 27: An Exploration of the Intrinsic Negative Socio-Economic Implications of ICT Interventions
Introduction
Socio-Economic Impact of ICT
Tools, Machines, and ICT
Routes of Elimination and Substitution
Conditions for Elimination and Substitution
Pragmatics and Ethics of Implementation
Dimensions of Social Impact of ICT
Platform, Message, and Target
Competing with Others: Additional Implications
Competing with Others: Social Implications
Impact of Collaboration
Investigating Negative Implications of ICT: What Is the Plan?
Conclusion
References
SECTION IV: Appendix X
The Purpose and the Suggested Use of the Content in this Appendix
Appendix X1: Models of Economic Growth
References
Appendix X2: A Model of the Socio-Economic Impact of ICT
References
Index
Recommend Papers

Quantitative Methodologies using Multi-Methods: Models for Social Science and Information Technology Research
 9780367903961, 9781032046976, 9781003024149

  • 0 0 0
  • Like this paper and download? You can publish your own PDF file online for free in a few minutes! Sign Up
File loading please wait...
Citation preview

Quantitative Methodologies using Multi-Methods

Quantitative Methodologies using Multi-Methods Models for Social Science and Information Technology Research

Sergey V. Samoilenko and Kweku-Muata Osei-Bryson

First edition published 2022 by Routledge 6000 Broken Sound Parkway NW, Suite 300, Boca Raton, FL 33487-2742 and by Routledge 2 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN © 2022 Taylor & Francis Group, LLC Routledge is an imprint of Taylor & Francis Group, LLC Reasonable efforts have been made to publish reliable data and information, but the author and publisher cannot assume responsibility for the validity of all materials or the consequences of their use. The authors and publishers have attempted to trace the copyright holders of all material reproduced in this publication and apologize to copyright holders if permission to publish in this form has not been obtained. If any copyright material has not been acknowledged please write and let us know so we may rectify in any future reprint. Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, microfilming, and recording, or in any information storage or retrieval system, without written permission from the publishers. For permission to photocopy or use material electronically from this work, access www.copyright.com or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 978750-8400. For works that are not available on CCC please contact [email protected] Trademark Notice: Product or corporate names may be trademarks or registered trademarks and are used only for identification and explanation without intent to infringe. Library of Congress Cataloging‑in‑Publication Data A catalog record for this title has been requested ISBN: 978-0-367-90396-1 (hbk) ISBN: 978-1-032-04697-6 (pbk) ISBN: 978-1-003-02414-9 (ebk) Typeset in Times by KnowledgeWorks Global Ltd.

Contents Preface: Possible Uses of this Book.......................................................................... xv Introduction.............................................................................................................xvii

Section I  D  evelopment of the Methodological Modules...................................................1 Chapter 1 Pre-Requisite General Questions.......................................................... 3 Impact of the Assumption of Homogeneity of the Sample on Research Questions.....................................................................4 From a Basket of Apples to a Set of Systems (Decision-Making Units)................................................................................................5 From Systems to Systems in Context....................................................6 Chapter 2 Components of Multi-Method Methodologies...................................... 9 Cluster Analysis (CA)............................................................................ 9 Classification Decision Trees Induction (CDTI)................................. 10 Neural Networks (NNs)....................................................................... 12 Association Rules Mining (ARM)...................................................... 14 Data Envelopment Analysis (DEA)..................................................... 16 Multiple Regression (MR)................................................................... 18 Chapter 3 Framework for Methodological Modules............................................ 21

Section II  D  escription of the Methodological Modules............................................... 27 Chapter 4 A1: Homogeneous Sample – DEA and DTI........................................ 29 Phase 1: DEA...................................................................................... 29 Phase 2: DTI........................................................................................ 30 Examples of Application of DEA and DTI......................................... 30 Chapter 5 A2: Homogeneous Sample – DEA and ARM..................................... 31

v

vi

Contents

Phase 1: DEA...................................................................................... 31 Phase 2: ARM..................................................................................... 31 Examples of Application of DEA and ARM...................................... 32 Chapter 6 B1: Heterogeneous Sample (Groupings Are Given) – DTI and ARM............................................................................................. 33 Phase 1: DTI........................................................................................ 33 Phase 2: ARM.....................................................................................34 Examples of Application of DTI and ARM........................................ 35 Chapter 7 B2: Heterogeneous Sample (Groupings Are Given) – DTI and MR................................................................................................ 37 Phase 1: DTI........................................................................................ 38 Option 1: DTI Using the Data Set Comprised of a Causal Model Only................................................................................ 38 Option 2: DTI Using the Data Set without Causal Model.............. 38 Option 3: DTI Using the Complete Data Set.................................. 38 Phase 2: MR........................................................................................ 38 Option 1: MR Using the Causal Model Only................................. 39 Option 2: MR Using the Adapted Causal Model – Contextual Independent Variable................................................................. 39 Option 3: Creating a New MR Using Contextual Independent Variables.................................................................................... 39 Example of Application of DTI and MR............................................ 39 Chapter 8 B3: Heterogeneous Sample (Groupings Are Given) – DTI, DEA, and ARM............................................................................................. 41 Phase 1: DTI........................................................................................ 42 Option 1: The Data Set Is Comprised of the Variables of the DEA Model................................................................................ 42 Option 2: The Data Set Contains Contextual Variables................. 42 Phase 2: DEA...................................................................................... 43 Phase 3: ARM..................................................................................... 43 Option 1: ARM to Generate “If→ (Level of the Top-Split Variable(s))”............................................................................... 45 Option 2: ARM to Generate “If→ (DEA Model’s Inputs)”........... 45 Option 3: ARM to Generate “If→ (DEA Model’ s Outputs)”....... 45 Option 4: ARM to Generate “If→ (Level of Averaged Relative Efficiency)”.................................................................. 45 Option 5: ARM to Generate “If→ (Received Categorization)”..... 45 Examples of Application of DTI, DEA, and ARM.............................46

Contents

vii

Chapter 9 B4: Heterogeneous Sample (Groupings Are Given) – DTI, DEA, and NN...................................................................................... 47 Phase 1: DTI........................................................................................ 48 Phase 2: DEA...................................................................................... 48 Phase 3: NN......................................................................................... 49 Step 1: Generate NN Model of Transformative Capacity..................................................................................... 49 Step 2: Generate Outputs of a Less Efficient Group Based on Transformative Capacity of a More Efficient Group............ 49 Step 3: Generate Outputs of a More Efficient Group Based on Transformative Capacity of a Less Efficient Group............. 49 Step 4: Compile the Generated Outputs in a New Data Set............................................................................. 50 Phase 4: DEA...................................................................................... 50 Example of Application of DTI, DEA, and NN.................................. 50 Chapter 10 C1: Heterogeneous Sample (Groupings Are Not Known) – CA and DTI............................................................................................... 51 Phase 1: CA......................................................................................... 51 Phase 2: DTI........................................................................................ 52 Examples of Application of CA and DTI............................................ 52 Chapter 11 C2: Heterogeneous Sample (Groupings Are Not Known) – CA and ARM............................................................................................. 55 Phase 1: CA......................................................................................... 55 Phase 2: ARM..................................................................................... 56 Option 1: ARM Using Only Intrinsic Variables............................. 56 Option 2: ARM Using Only Contextual Variables........................ 56 Option 3: ARM Using Intrinsic and Contextual Variables.................................................................................... 57 Examples of Application of CA and ARM......................................... 57 Chapter 12 C3: Heterogeneous Sample (Groupings Are Not Known) – CA, DTI, and MR....................................................................................... 59 Phase 1: CA.........................................................................................60 Phase 2: DTI........................................................................................60 Option 1: Data Set Is Limited to Variables of the MR Model..................................................................................60 Option 2: Data Set Comprises Variables of the MR Model and Contextual Variables...........................................................60 Phase 3: MR........................................................................................ 61 Example of Application of CA, DTI, and MR.................................... 62

viii

Contents

Chapter 13 C4: Heterogeneous Sample (Groupings Are Not Known) – CA, DTI, and ARM.................................................................................... 63 Phase 1: CA.........................................................................................64 Phase 2: DTI........................................................................................64 Option 1: A Priori Target Variable................................................. 65 Option 2: CA-based Target Variable.............................................. 65 Phase 3: ARM..................................................................................... 65 Step 1..............................................................................................66 Step 2..............................................................................................66 Step 3..............................................................................................66 Step 4..............................................................................................66 Examples of Application of CA, DTI, and ARM............................... 67 Chapter 14 C5: Heterogeneous Sample (Groupings Are Not Known) – CA and DEA.............................................................................................. 69 Phase 1: CA......................................................................................... 69 Option 1: CA based on the DEA Model......................................... 69 Option 2: CA based on the DEA Model and Contextual Variables.................................................................................... 70 Phase 2: DEA...................................................................................... 71 Examples of Application of CA and DEA.......................................... 71 Chapter 15 C6: Heterogeneous Sample (Groupings Are Not Known) – CA, DEA, and ARM.................................................................................. 73 Phase 1: CA......................................................................................... 73 Phase 2: DEA...................................................................................... 74 Phase 3: ARM..................................................................................... 75 Option 1: Complete Sample, # of Variables = the DEA Model..... 75 Option 2: Complete Sample, # of Variables = the DEA Model + Contextual Variables................................................... 75 Option 3: Sub-Sets of the Sample, # of Variables = the DEA Model......................................................................................... 76 Option 4: Sub-sets of the Sample, # of Variables = the DEA Model + Contextual Variables................................................... 76 Examples of Application of CA, DEA, and ARM.............................. 76 Chapter 16 C7: Heterogeneous Sample (Groupings Are Not Known) – CA, DTI, and DEA..................................................................................... 77 Phase 1: CA......................................................................................... 77 Phase 2: DTI........................................................................................ 78 Phase 3: DEA...................................................................................... 78 Examples of Application of CA, DTI, and DEA................................ 79

Contents

ix

Chapter 17 C8: Heterogeneous Sample (Groupings Not Known) – CA, DTI, DEA, and NN...................................................................................... 81 Phase 1: CA......................................................................................... 82 Phase 2: DTI........................................................................................ 83 Phase 3: DEA...................................................................................... 83 Phase 4: NN.........................................................................................84 Step 1: Creating an NN Model of “Low-Level” Cluster........................................................................................84 Step 2: Creating an NN Model of “High-Level” Cluster........................................................................................84 Step 3: Simulation of the Outputs of “Low-Level” Cluster Using NN Model of “High-Level” Cluster................................84 Step 4: Simulation of the Outputs of “High-Level” Cluster Using NN Model of “Low-Level” Cluster................................. 85 Phase 5: DEA...................................................................................... 85 Examples of Application of CA, DTI, DEA, and NN......................... 86

Section III  M  ethodological Modules – Examples of Their Application............................. 87 Chapter 18 A Hybrid DEA/DM-based DSS for Productivity-Driven Environments...................................................................................... 89 Introduction......................................................................................... 89 Description of the DSS........................................................................90 Externally Oriented Functionality..................................................90 Internally Oriented Functionality...................................................92 Architecture of the DSS................................................................. 93 An Illustrative Application..................................................................94 Step 1: Is the Business Environment Homogeneous?..................... 95 Step 2: What Are the Factors Responsible for Heterogeneity of the Business Environment?...................................................96 Step 3: Do Groups of Competitors Differ in Terms of the Relative Efficiency?...................................................................97 Step 4: What Are some of the Factors Associated with the Differences in Relative Efficiency?...........................................99 Step 5: Are There any Complementarities Between the Relevant Variables?.................................................................. 101 Step 6: What Is a Better Way to Improve Production of Outputs?................................................................................... 102 Conclusion......................................................................................... 103 Acknowledgment............................................................................... 104 References......................................................................................... 104

x

Contents

Chapter 19 Determining Sources of Relative Inefficiency in Heterogeneous Samples: Methodology Using Cluster Analysis, DEA, and Neural Networks............................................................................... 109 Introduction....................................................................................... 109 Description of the Methodology....................................................... 111 Description of Steps 3–5 of the Methodology.............................. 113 Step 3: Generate a “Black Box” Model of Transformative Capacity of Each Cluster.................................................... 113 Step 4: Generate Simulated Sets of the Outputs for Each Cluster................................................................................. 113 Step 5: Determine the Sources of the Relative Inefficiency of the DMUs in the Sample................................................ 113 Motivation for Steps 3 and 5 of the Methodology........................ 113 Motivation for Step 3............................................................... 113 Motivation for Step 5............................................................... 114 Illustrative Example.......................................................................... 115 Description of the Illustrative Data Set........................................ 115 Application of the Methodology on the Illustrative Data Set....... 116 Results of Step 1: Evaluate the Scale Heterogeneity Status of the Data Set..................................................................... 116 Results of Step 2: Determine the Relative Efficiency Status of Each DMU........................................................... 117 Results of Steps 3 and 4: Generate Simulated Sets of the Outputs for Each Cluster Based on Black Box Models Transformative Capacity Processes.................................... 118 Results of Step 5...................................................................... 118 Discussion and Conclusion................................................................ 119 Acknowledgment............................................................................... 121 References......................................................................................... 121 Chapter 20 Exploring Context Specific Micro-Economic Impacts of ICT Capabilities........................................................................................ 125 Introduction....................................................................................... 125 Theoretical Framework and the Research Model............................. 128 The Methodology of the Study.......................................................... 131 Phase 1: Application of Data Envelopment Analysis (DEA)....... 131 Phase 1, Step 1......................................................................... 132 Phase 1, Step 2......................................................................... 132 Phase 1, Step 3......................................................................... 132 Phase 2: Decision Tree-Based Analysis....................................... 133 Phase 2, Step 1......................................................................... 133 Phase 2, Step 2......................................................................... 134 Description of the Data..................................................................... 134 Results of the Data Analysis............................................................. 134

Contents

xi

Results from Phase 1: Application of Data Envelopment Analysis (DEA)........................................................................ 134 Phase 1, Step 1......................................................................... 134 Phase 1, Step 2......................................................................... 135 Phase 1, Step 3......................................................................... 136 Results from Phase 2 – Decision Tree (DT) Based Analysis................................................................................... 137 Conclusion......................................................................................... 139 Contributions to Theory.................................................................... 140 Contributions to Practice................................................................... 141 Acknowledgment............................................................................... 141 References......................................................................................... 141 Chapter 21 A Methodology for Identifying Sources of Disparities in the Socio-Economic Outcomes of ICT Capabilities in SSAs.............................................................................................. 143 Introduction....................................................................................... 143 Research Framework......................................................................... 145 Proposed Methodology..................................................................... 146 A New Methodology: Benefits and Justifications......................... 146 Phase 1: Data Envelopment Analysis (DEA)................................ 147 Phase 2: Decision Tree Induction (DTI)...................................... 148 Phase 3: Association Rule Mining (ARM).................................. 148 Research Questions and Null Hypotheses of the Study.................... 148 The Data............................................................................................ 149 Results of the Data Analysis............................................................. 150 Phase 1: Data Envelopment Analysis........................................... 150 Phase 2: Decision Tree Induction................................................. 152 Phase 3: Association Rule Mining............................................... 153 Discussion of the Results................................................................... 154 Conclusion......................................................................................... 155 Acknowledgment............................................................................... 155 References......................................................................................... 155 Chapter 22 Discovering Common Causal Structures that Describe ContextDiverse Heterogeneous Groups......................................................... 157 Introduction....................................................................................... 157 A Conceptualization of the Benchmarking Problem........................ 158 Research Problem and Research Questions of the Study.................. 159 The Proposed Methodology.............................................................. 168 Description of the Methodology................................................... 168 Justification & Benefits of the Methodology................................ 169 Illustrative Example – Application to Sub-Saharan Economies....... 172 Phase 1: Define the Transformation Framework.......................... 172

xii

Contents

Phase 2: Partition the Set of Decision Making Units into Meaningful Groups................................................................. 173 Phase 3: Data Envelopment Analysis........................................... 173 Phase 4: Decision Tree Induction (DTI)...................................... 176 Phase 5: Association Rule Mining............................................... 177 Conclusion......................................................................................... 178 Acknowledgment............................................................................... 179 References......................................................................................... 179 Chapter 23 An Empirical Investigation of ICT Capabilities and the Cost of Business Start-up Procedures in Sub-Saharan African Economies......................................................................................... 183 The Research Framework and Research Questions.......................... 186 Proposed Methodology..................................................................... 187 Phase 1: Cluster Analysis (CA)..................................................... 187 Phase 2: Decision Tree Induction................................................. 188 Phase 3: Data Envelopment Analysis........................................... 188 Phase 4: Ordinary Least Squares Regression............................... 188 Phase 5: Association Rule Mining............................................... 188 Data................................................................................................... 188 Results of the Data Analysis............................................................. 189 Phase 1: CA.................................................................................. 189 Phase 2: DTI................................................................................. 189 Phase 3: DEA................................................................................ 191 Phase 4: OLS................................................................................ 191 Phase 5: ARM.............................................................................. 192 Interpretation of the Results of the Data Analysis............................ 194 Cluster Analysis............................................................................ 194 Decision Tree Induction............................................................... 195 DEA.............................................................................................. 196 Ordinary Least Squares (OLS)..................................................... 197 ARM............................................................................................. 197 Discussion of the Results of the Study.............................................. 198 Conclusion.........................................................................................200 Acknowledgment............................................................................... 201 References.........................................................................................202 Chapter 24 Exploring the Socio-Economic Impacts of ICT-Enabled Public Value in Sub-Saharan Africa............................................................ 203 Introduction....................................................................................... 203 Research Framework of the Study....................................................204 Research Questions of the Study.......................................................207 Methodology of the Investigation......................................................208 Phase 1: Cluster Analysis.............................................................209

Contents

xiii

Phase 2: Decision Tree Induction.................................................209 Phase 3: Data Envelopment Analysis (DEA)...............................209 Phase 4: Ordinary Least Squares (OLS) Regression....................209 Phase 5: Association Rule Mining (ARM)...................................209 Data...................................................................................................209 Results of the Data Analysis............................................................. 210 Phase 1: CA.................................................................................. 210 Phase 2: DTI................................................................................. 211 Phase 3: DEA................................................................................ 212 Phase 4: OLS................................................................................ 212 Phase 5: ARM.............................................................................. 214 Discussion of the Results of the Study.............................................. 216 Conclusion......................................................................................... 218 Acknowledgment............................................................................... 218 References......................................................................................... 218 Chapter 25 Contributing Factors to Information Technology Investment Utilization in Transition Economies: An Empirical Investigation.... 221 Introduction....................................................................................... 221 Theoretical Framework.....................................................................224 Growth Accounting......................................................................224 Theory of Complementarity......................................................... 225 Overview on the Data........................................................................ 227 Methodology: Searching for the Determinants of the Efficiency of Utilization of Investments in Telecoms.................................... 228 Phase 1: Data Envelopment Analysis........................................... 228 Data Used to Perform DEA..................................................... 228 Phase 2: Cluster Analysis............................................................. 229 Data Used to Perform CA........................................................ 230 Phase 3: Decision Tree................................................................. 230 Data Used to Perform DT........................................................ 230 Results............................................................................................... 231 Results: DEA................................................................................ 231 Results: Cluster Analysis.............................................................. 232 Results: Decision Tree.................................................................. 234 Contribution of the Study.................................................................. 238 Summary and Conclusion................................................................. 239 Acknowledgment...............................................................................240 References.........................................................................................240 Appendix A....................................................................................... 243 Chapter 26 Increasing the Discriminatory Power of DEA in the Presence of the Sample Heterogeneity with Cluster Analysis and Decision Trees..... 245 Introduction....................................................................................... 245

xiv

Contents

The Proposed Methodology.............................................................. 247 Overview of Data Set of Illustrative Example.............................. 247 Description of the Methodology...................................................248 Step 1: Determine the Structural Homogeneity Status of the Data Set......................................................................... 249 Step 2: Determine the Relative Efficiency Status of DMUs..... 251 Step 3: Describe the Relative Efficiency Categories............... 253 Conclusion......................................................................................... 255 Acknowledgment............................................................................... 258 References......................................................................................... 258 Chapter 27 An Exploration of the Intrinsic Negative Socio-Economic Implications of ICT Interventions..................................................... 259 Introduction....................................................................................... 259 Socio-Economic Impact of ICT........................................................260 Tools, Machines, and ICT............................................................260 Routes of Elimination and Substitution........................................ 261 Conditions for Elimination and Substitution................................ 262 Pragmatics and Ethics of Implementation.................................... 263 Dimensions of Social Impact of ICT............................................264 Platform, Message, and Target.....................................................264 Competing with Others: Additional Implications........................ 267 Competing with Others: Social Implications............................... 268 Impact of Collaboration................................................................ 268 Investigating Negative Implications of ICT: What Is the Plan?....... 269 Conclusion......................................................................................... 271 References......................................................................................... 271

Section IV  Appendix X........................................ 273 The Purpose and the Suggested Use of the Content in this Appendix............................................................................... 273 Appendix X1 Models of Economic Growth......................................................... 275 References......................................................................................... 278 Appendix X2 A Model of the Socio-Economic Impact of ICT............................ 279 References......................................................................................... 282 Index....................................................................................................................... 283

Preface: Possible Uses of this Book It is only expected that our reader should be presented with a possible explanation, before actually reading this book, for why this text should be read at all. Sir Winston Churchill once quipped that he only read for pleasure or profit, hence, following this binary option exercised by this well-regarded individual, we suggest that it is a profit that our reader could derive from reading the material we present in this book. Specifically, we suggest that our work could be used for five different, but not mutually exclusive, purposes. First, our reader may approach the book as one of the means for a user-friendly introduction to the popular methods of data mining and data analysis. In our presentation and overviews of the methods, we avoided getting involved into details that are more suitable for more advanced users – we assume that our audience has, at most, surface-level knowledge of the methods presented in the book. The rationale was that once a reader gets an idea, a conceptual grasp, regarding how a method works, then she will be able to get a more detailed overview from other sources. After all, it is not an exaggeration that a separate book could be written (and they are written) in order to give justice to each of the methods presented in our text. Second, given a reader's introductory knowledge of the methods covered in this text, our book could serve as an introductory guide to the subject of complementarity of the tools and techniques of data analysis – we aim to demonstrate how the methods could be used in synergy in order to offer insights into the issues that could not be dissected by any single method alone. We do not suggest that the methodological modules we describe in the book are the only combinations possible. Instead, we advise our reader to create her own modules, while keeping in mind a necessity of justifying the selection of methods and arguing the beneficial nature of their interaction to her own audience. After all, the resultant methodology has to make sense not only to the investigator, but also to the audience that is interested in the subject. On a side note: we found that the structure of the type “Assumption/Justification” works well for the purpose of making any assumptions and assertions explicit and for eliciting a rational explanation for why the selected methodology is one that is worthy of consideration. Third, we suggest that this text could be used as a set of templates, where, given a set of research questions, the investigator could identify a set of methodological modules allowing for answering the research questions of interest. This is not entirely unlike the relationship between analysis and design phases of the systems development life cycle – where the “What?” of the analysis phase has to be translated into the “How” of the design phase. So, we hope that our reader, with the help of this book, will be able to identify methodological modules (the “How”) that are suitable for answering her research questions (the “What”). Simply put, we hope that our work would help in transitioning a conceptual domain of the research questions into xv

xvi

Preface: Possible Uses of this Book

a scaffolding of the data analytic and data mining methods by which the research questions could be answered. Also, this text could serve as a guide to exploring what the data holds. Let us say, a researcher is in possession of a set of data, complemented with domain area’ knowledge. Usually, there are frameworks and theories that could be used to conduct scientific inquiries, but, sometimes, it is not clear what sort of questions the data may answer. In such case (which is somewhat similar to grounded research-type of an approach), the investigator may use the methodological modules presented in this book (corresponded with the running examples) to generate a set of preliminary questions which, after a careful consideration and a requisite culling, could be formulated into a set of questions consistent within a selected theory or a framework. Finally, we suggest that our work could be used as a generator of new research questions. This is because, as a reader would have an opportunity to discover, the result of applying every method in each of the modules opens up a new dimension ripe with follow-up questions of the type “Why is this so?” Consequently, even if an inquiry started with a set of research questions based on a theory or a framework, at the later point the study could become augmented by the additional follow-up questions that arose as a result of the application of the selected methodology.

Introduction An investigator undertaking a research project usually faces a variety of challenges to overcome, the following four being the most obvious – we suggest them without implying that their order is an indicator of a relative significance. The first challenge is associated with the expression of the overall goal of the study, corresponded by the formulation of a set of well-defined research question(s). The second challenge is that of selecting the underlying framework of the study, where the level of granularity of the research should be supported by the appropriate level of precision, or detail, of the research framework (e.g., level of grand theory vs. level of a case study). The third challenge is the availability of the data. Finally, there is a challenge of selecting or designing an appropriate methodology that allows for analyzing the available data and answering the research questions of the study. We have dealt with challenges of creating theoretical frameworks in our prior work (e.g., Samoilenko & Osei-Bryson, 2017), and now we concentrate our effort on the methodological aspect of data analytic inquiry. Commonly, it is a case that the decision regarding the appropriate methodology of the study, as well as the decision concerning the selection of the data analytic methods, is impacted by the nature of the available data. It is fair to say that the question of the quality and availability of the data is unique to each study. However, the general guidelines are straightforward – get as much of the relevant good quality data as possible. In some cases, the primary data must be collected by the investigators, while in other cases the secondary data could be obtained from reliable public sources. In this text our aim is to illustrate how quantitative, multi-method methodologies could be applied to answer a variety of research question. As a result, our intent is to offer our audience a general map, in a form of a methodological framework, that could be applied to a variety of research scenarios with the purpose of crafting and answering research question that are fairly typical of quantitative inquiries based on the analysis of non-parametric data. We intend to achieve our goal by demonstrating a creation of what we call methodological modules – synergistic combinations of two or more data analytic techniques that could be used independently, or combined together to create complex methodologies. We are aware that the investigators need to defend the validity of the methodology of the study to their audience and the reviewers, and in order to assist our reader in this regard we offer some published examples and their citations for each methodological module we describe. Samoilenko, S. and Osei-Bryson, K.-M. (2017). Creating Theoretical Research Frameworks Using Multiple Methods: Insight from ICT4D Investigations. Auerbach Publications. ISBN-13: 978-1498779951.

xvii

Section I Development of the Methodological Modules

1

Pre-Requisite General Questions

In crafting a multi-method methodology, a “slice of an apple” analogy is applicable – each method is akin to a knife slicing an apple in its own way under a different angle. The greater the variety of slices, the richer the perspective we can obtain on how an apple looks inside. The question to answer is, of course, how to arrange the process of slicing in a complementary way, especially if we are dealing with a basket of apples, and, furthermore, with a basket of apples of possibly different kinds. Let us consider our “basket of apples” analogy and consider some of the questions that could be asked. Do we assume in our investigation that the apples in the basket are of the same kind? The positive answer to this question reflects the assumption of homogeneity of the sample. Consequently, we are looking, conceptually, for a single apple that is representative of the whole basket. And, of course, we don’t assume that our basket of apples is just like all other basket of apples out there in this world – we don’t generalize our findings beyond the basket. Do we assume that the apples in the basket are of different kinds? The basic assumption stemming from a positive answer to this question is that of heterogeneity of the sample. Resultantly, the obvious question to pose will be regarding the number of different kinds of apples in our basket. This is something that needs to be discovered, for we don’t have any “received” knowledge neither regarding the categories of apples in the basket, nor regarding the number of the categories. Do we know that the apples in our basket are of different kinds? The answer to this question comes in the form of the received knowledge expressed in the form of categories defining heterogeneous sub-sets of the apples in our basket. The important point to make here is that the set of categories is not necessarily definitive and final; instead, it is simply a categorization based on the value of a chosen by the external agency’ attribute. For example, we can be told that the basket of apples consists of the sub-sets based on color. However, we may also be told that the sub-sets differ in terms of size. Or, in terms of taste (e.g., sweet, tart, and sour), or, in terms of anything else that can be usefully applied for the purposes of partitioning the sample (e.g., perceived attractiveness and freshness). 3

4

Quantitative Methodologies using Multi-Methods

IMPACT OF THE ASSUMPTION OF HOMOGENEITY OF THE SAMPLE ON RESEARCH QUESTIONS The questions above could be considered to be a “first order” questions, and after answering them we could proceed to generating much more interesting, “higher order” questions. Clearly, the most limited in terms of the follow-up questions is the option based on the assumption of homogeneity of the sample. Fundamentally, our “higher order” questions are fairly simple, and are of descriptive-predictive nature. For example, we could, assuming the homogeneity of the sample, ask the following question: What is a common set of attributes and their values describing the basket of apples? Additionally, given a time series data, we could pose a question such as: What are some of the changes that took place over time that impacted the apples in the basket? This, fundamentally, would allow us to pose a pretty much the last interesting question of the scenario, namely: Given the changes that impacted our basket of apples over time, what are some of the changes that will impact our apples in the future? Given the scenario described above, it is easy to see that answering such question does not require a multi-method methodology; instead, a single method will do just fine. Unlike the assumption of homogeneity, however, the assumption of heterogeneity of the sample gives rise to a much richer set of “higher order” questions. For example, in the absence of the received knowledge regarding the categories, an investigator may ask the following question: What are some of the sub-groups that are present in the sample? A “basket of apples” scenario may offer an answer of the type: The basket of apples consists of three groups of apples – red, yellow, and green. The very logical follow-up to such questions is the question regarding the nature of the differences of the sub-groups, namely: Given the presence of multiple sub-groups in the sample, what are some of the factors responsible for the heterogeneity? Keeping in mind our example, we can come up with such answer, as: an apple variety is responsible for the membership in a group. Or, the levels of chlorophyll and carotenoids impact the apple’ group membership.

Pre-Requisite General Questions

5

Similarly, in the presence of the received knowledge regarding the differentiation of the sub-groups, a researcher may ask a question of the type: Given the sub-sets of the sample based on the received criterion, what other factors/attributes are associated with differentiating criterion? Again, in addition to the difference in the level of pigments that determine the group membership, an investigator may discover that the levels of exposure to sun and of the temperature also do impact the color of apples. This may lead to an inquiry into the factors that differentiate the sub-groups the most, as well as the least. As we can see the assumption of heterogeneity of the sample gives rise to a set of the related research questions, where each question will require an application of its own data analytic method to answer. At this point we can state a very simple rule of thumb: an assumption of heterogeneity of the sample gives rise to the application of multi-method methodologies.

FROM A BASKET OF APPLES TO A SET OF SYSTEMS (DECISION-MAKING UNITS) A much more sophisticated set of “second order” questions could be generated if we consider our sample to be comprised not of a basket of apples, but of a basket of systems. For our intents and purposes, we define a system as a structurally and functionally complex entity that receives inputs and transforms them into outputs. And, of course, systems that are of interest to us are open, dynamic, non-linear, and complex. For all intents and purposes economies, departments, hospitals, universities, NBA players, assembly line workers are all examples of systems that could be represented/referred to as units that transform inputs into outputs. This gives us another three dimensions to consider, namely, level of inputs, level of outputs, and efficiency of conversion of inputs into outputs. It goes without saying that “input” and “output” are logical designations – a system could have multiple inputs and multiple outputs. Under the assumption of homogeneity of the sample we can generate, for all intents and purposes, three types of “second order” questions, for example: What is the average (or, min, max, etc.) level of inputs? What is the average (or, min, max, 25th percentile, and so on) level of outputs? Additionally, this basic model describing the system from the perspective of inputs, outputs, and the process of transformation opens up an access to the questions associated with the relative performance of the system. This leads to the questions of the type of: What is the level of performance of the system – the level of average (smallest, greatest) level of transformative capacity (conversion of inputs into outputs)?

6

Quantitative Methodologies using Multi-Methods

It is easy to see, in the case of the assumption of homogeneity of the sample, that such questions could be answered via using a very simple methodology – a single descriptive statistics-type method will do the job. The methodological simplicity disappears once we operate under the assumption of the heterogeneity of the sample. In addition to the mentioned above three questions, which will need to be answered per sub-group, we can generate such additional questions as: What are the differences in inputs/outputs/transformative capacity between groups? What are the sources of the differences? Is the difference in outputs is due to the difference in inputs, or is it due to the difference in efficiency of the process of transformation? What are the reasons for the differences in transformative capacity? If the investigator is in the possession of the time series data, then another set of questions related to the performance of the system over time could be generated, such as: What are the changes in the level of performance of the system over time? Once the changes have been identified, an investigator may be interested to find out the answer to the follow up question, namely What are some of the reasons for changes in performance of the system over time? Intuitively, it is easy to see that answering “second order” questions will require employing a methodology that is not based on one or two data analytic methods. However, the complexity of a suitable methodology increases even more if we consider systems in their contexts.

FROM SYSTEMS TO SYSTEMS IN CONTEXT The consideration of a context of the systems – decision-making units – gives rise to even more sophisticated set of, what could be called “third order”, questions. Such inquiries, fundamentally, deal with investigating the impact of the environment on the relative level of performance of the system and could be generalized in the form of a following question: What is the impact of the environmental factors on the level of the relative performance of the system? The basic premise behind such question is that the environment impacts the performance of a set of heterogeneous systems differently. And, of course, this leads to a significant increase in the number of attributes describing the system, for now we need to consider not only inputs, outputs, and transformative capacity but also the factors that could plausibly influence them from outside.

Pre-Requisite General Questions

7

A simple illustration is provided by a comparison of various types of engines – any and all engines require some sort of a fuel, produce some form of energy that is applied to do work, and require a transformative capacity to convert fuel into energy. It is easy to see that temperature, humidity, air pressure, as well as other factors, have a capacity to impact fuel, the process of conversion of fuel into energy, as well as relative amount of work produced by the generated energy. Resultantly, an investigator could ask the following “third order” questions: What environmental factors influence inputs of the system? What environmental factors influence outputs of the system? What environmental factors influence the transformative capacity of the system? The above mentioned questions and the scenario applies to the case when we consider the environment of our sample of units to be homogeneous. We can ask even more interesting questions if operate under the assumption that our systems reside in different environments – when we are dealing with a heterogeneous sample comprised of sub-groups, where each subgroup resides in its own specific context. Let us consider a sample consisting of retail stores – the previous scenario would allow us to compare the stores located in approximately the same climate and operating within similar socio-economic areas/neighborhoods. However, what if we decide to compare the levels of relative performance of an athletic shoe store located in California with a work boots store operating in Alaska? Given this scenario, an investigator will need to find out the environmental factors common to both stores. It is important to note that by “environmental factors”, we denote a wide variety of those factors that are external to the common “input-output” model describing the stores. Consequently, a researcher investigating the factors influencing the difference in the performance of the systems/units functioning in different contexts may ask a following question: Do heterogeneous contexts of the sub-groups in the sample share common factors? Once this question is answered, the follow-up question begs to be asked, namely: Do common environmental factors have a similar impact on the level of the relative performance of the systems functioning in heterogeneous contexts? It is easy to see, by now, that an inquiry addressing such question would have to rely on a methodology comprised of multiple methods of data mining and data analysis.

2

Components of MultiMethod Methodologies

At this point, we would like to provide our readers with a very brief and informal overview of the data analytic tools that will be used to create what we call “methodological modules” – logically related collections of methods that could be applied in order to answer a variety of complex research questions. The following context is not intended to be a thorough introduction, but, rather, a jargon-free and novice-friendly introduction that should be followed by a more comprehensive coverage we offered in our previous work.

CLUSTER ANALYSIS (CA) The purpose of CA is to test the data set for the presence of heterogeneous subgroups. So, if we have a sample of National Hockey League (NHL) players, then CA may offer us insights regarding the similarity of the hockey players in our sample. Of course, it is always possible to assume that the data set consists of homogeneous entities, that is, if one uses criteria allowing for such assumption. For example, a set of NHL players could be considered a homogeneous based on the fact that all of the members of the set are, indeed, members of NHL teams, or, they are all hockey players, or, because they all have been drafted and given a draft number. But, most of the time it is useful to search for characteristics of hockey players based on more granular differentiation. For example, the NHL players may differ in regard to the position they play, based on the years of experience, based on height and weight, or any multitude of factors that could be of interest to the investigator. However, the problem with creating sub-sets of the sample is based on the choice of criteria, for, in the case of the self-selected criterion, a researcher would almost always face a reviewer’s question: Why did you select this criterion? CA, being an unguided technique, helps in identifying sub-sets of the sample in a fashion that is somewhat independent of the researcher. An “unguided technique” means that CA partitions the sample based on the criteria of the method itself – CA will attempt to identify the presence of the clusters in such way that the members of each cluster are similar to each other, but the clusters are different from each other. As a result, we are getting groups that are “similar inside and different outside”. It is always possible, of course, that CA will not produce any clusters – this will happen if the data set is comprised of truly homogeneous entities. The “truly” part, however, carries a connotation of subjectivity – who determines what “truly” means? After all, it is easy to make an argument that the differences are there, but an investigator simply assumed their absence. And this is where an investigator influences the results of CA – via the selection of settings according to which clustering will be performed. Let us consider the most popular types of CA. 9

10

Quantitative Methodologies using Multi-Methods

K-means clustering approaches the data set with the requirement to identify K-sub-groups within the sample, where K is chosen by the investigator number. For example, if we have 100 NHL players in our sample, then the selection of K = 3 will result in three sub-groups, the selection of K = 2 gives us two, and of K = 4 will give us four sub-groups of hockey players. Divisive hierarchical clustering approach is based, unsurprisingly, on dividing, in a top-down fashion, a complete set into a number of clusters, where the stopping criterion is set by an investigator. So, the starting point of this approach is a complete data set that is being partitioned in a step-by-step approach until a specified criterion is met. Agglomerative hierarchical clustering, on the other hand, starts with the consideration of every member of the set to be, so to speak, its own cluster, and proceeds in a bottom-up fashion, to combine – to agglomerate – the members of the set into a larger set of clusters. Both hierarchical approaches operate in a step-by-step fashion, and the investigator is given an opportunity to see the intermediate results of each step – one can actually see the process by which the clusters are formed. Regardless of the choice of the clustering method, the result of CA is a set of sub-groups – clusters – comprising the sample. So, it is easy to see that while the investigator has an influence on the number of clusters, the basis for clustering is method-driven and unguided – the selection of the variable based on which CA is performed is not under the control of the investigator. Meaning, if the results of CA yielded three clusters, then we know that the sample is comprised of three groups, but we don’t know, really, what factor or factors are responsible for the heterogeneity of the sample. An important point to consider is the relative size of clusters. Let us say, our sample of 100 NHL players was clustered into three groups. What should be a minimum size of the smallest group? Would the memberships such as Cluster 1 = 80, Cluster 2 = 17, and Cluster 3 = 3 be useful in the investigation? This is not an easy question to answer. Clearly, Cluster 1 and Cluster 2 are useful to consider, but Cluster 3 looks like an outlier. In our own research and our previous work, we suggested a rule of thumb for the smallest cluster to be at least 10% of the sample. So, the results of CA yield a set of clusters comprising the data set and allow for answering the question of relative homogeneity of the sample. But, the results of CA do not allow for answering the question of what is the underlying reason for the differences. The next method, Classification Trees, allows for gaining an insight into the nature of the difference between the clusters.

CLASSIFICATION DECISION TREES INDUCTION (CDTI) There are two general reasons for why an investigator may want to use decision tree induction. One of the reasons is to predict the result of something – of a value measured on interval scale. For example, given a previously constructed model of the relationships between inputs and outputs of an NHL player, we would be able to predict a productivity (e.g., output) of the incoming player based on the initial conditions (e.g., minutes played per game, experience in the league, etc.). The second reason is to classify an entity according to the pre-defined categories. In this overview, we

Components of Multi-Method Methodologies

11

concentrate on a decision tree’s analysis with the purposes of classification – CDTI. For example, given a general model of an NHL player, based on the categorization of “star”, “mid-level”, and “flop”, we will be able to place a given player within one of the pre-defined categories. But the most important use of CDTI is to uncover the reasons for why the categories differ from each other. Let us consider, again, a sample of NHL players that was partitioned into the “star”, “mid-level”, and “flop” categories. It is important to note that the reason for the partitioning is not important – this could be due to the received knowledge, or this could be due to the results of our own inquiry (e.g., cluster analysis). For example, we could also create groupings based on the position of the players – “defense” vs. “forwards” vs. “goaltenders”, or based on the geographic area of the player’s origin – “North America” vs. “Eurasia”. Or, the sub-sets could be created based on such criteria as “number of year in the league” or “number of games played”, or “time spent on injured list” and so on. In the case of our example, grouping the players into “star”, “mid-level”, and “flop” categories gives us an opportunity to place a player within one of the three categories. However, the reason for why the categorization was made is not clear – we, really, don’t know what separates “mid-level” from “flop”, we simply know that a player is in one group, or in another. But what we really would like to find out is the difference between the groups – after all, the difference could be bridgeable in a given case, and if we could help transitioning “flops” to “mid-level”, so much the better. By creating a target variable, let us say, “Group”, and assigning one value per group – “1” for “star”, “2” for “mid-level”, and “3” for “flop” – we can include this variable in our data set and run CDTI. So, if we had 20 variables describing NHL players, we will end up with 20 + 1 = 21 variables. Once we designated the “Group” variable as the “target” variable, CDTI will attempt to construct – to “grow” – a decision tree. The decision tree will take the form of an upside-down model of a real tree, with the roots and a trunk representing a complete data set (e.g., 21 variables multiplied by whatever number of NHL players we got in the sample) that is being, in a step-by-step fashion, partitioned into branches and then leaves. Once we generated the decision tree, there are two things of immediate interest (again, we remind our reader that in this overview we concentrate on the basics only), namely, what are top-level splits and what are the memberships of the toplevel nodes. The top-level splits are important because CDTI “grows” the trees on the basis of the variables differentiating the nodes at each level the most. Thus, the top splits will be based on the variables that are the most important in splitting the data set into the groups. In the case of our NHL example, the top-level splits could be based on scoring points (a sum of goals and assists) and time on ice (a total time on ice in the current season). The knowledge of the two most important variables, as well as of the values, levels, of the variables producing the split helps us to identify the most important criteria differentiating “star” from “mid-level” from “flop”. It is also helpful to consider that the generated decision tree could be presented not only in its graphical format but also in the format of English rules, which is a textual description of the tree. However, the practical and interpretive significance of the top-level split variables is impacted by the nodes’ membership.

12

Quantitative Methodologies using Multi-Methods

The relative membership of the top-level nodes – the sub-groups of the sample produced by the top-level splits – tells us how “clean” the separation between groups is. Meaning, we would like to get a series of splits placing our groups of hockey players neatly into their assigned categories. And, we can see that if we use scoring points and time on ice as the top-level split variables, then they should produce very clean separation between the groups based on the levels of the respective values. For example, we would reasonably expect that “star” will have high values in terms of scoring points and time on ice, “flop” will have low values, and “mid-level” will reside somewhere in between. However, let us imagine that we don’t have in our possession a set of variables that are designed to differentiate the groups – we don’t have stats to separate NHL players into the performance-based categories. Consider utilizing such attributes as height, weight, years of experience, and age. Then, we could easily see the results of CDTI being much less clear – we will see, if the top-level splits are made based on height and weight, a much more mixed picture, where each node of the decision tree will be populated by the mixture of “star”, “mid-level”, and “flop” representatives.

NEURAL NETWORKS (NNs) A NN is another technique that is often used for the purposes of prediction, but, unlike decision trees induction (DTI), a NN allows for predicting values of multiple variables. There are two basic requirements for using a NN: First, the data set must be comprised of interval data, and, second, the data set must be relatively large. Structurally, a NN could be described as a system comprising three parts: A set of input nodes, a set of output nodes, and a black box in between. The contents of the black box are comprised of the collection of one or more layers of inner, or hidden, nodes. The purpose of input nodes is to provide a set of values that are considered to be not necessarily input values to a process of transformation, but, rather, represent a set of starting conditions. For example, in the case of NHL players, we could specify such inputs as height, weight, age, years of experience as inputs. Here, we can see that such attribute as height is not an input to any process of transformation, but it is simply a one of the “givens” – something that a decision maker could designate as one of the points of departure. The purpose of output nodes is to specify, similarly, not the result of transformation of inputs via the transformative capacity of the system, but the end result of something of interest. In our example of hockey players, output nodes could be represented by the minutes played, goals, assists, and so on. Clearly, one would be hard pressed to explain, for example, the process that produces such value as “minutes played” for a given hockey player. The sole purpose of the hidden nodes is to connect, via a “black box” model, a set of input nodes to the set of output nodes. Structurally, hidden nodes are represented by weights that are derived/calculated, or obtained, during the process of “training” of a NN. Functionally, hidden nodes serve the purpose of translating inputs into outputs – not shedding light on the process of transformation, not explaining, not

Components of Multi-Method Methodologies

13

revealing anything regarding the mechanism of transformation, but simply supplying a collection of weights allowing for converting inputs into outputs. Let us consider an example of hockey players with the input nodes comprising {height, weight, age, years of experience in NHL} and the output nodes being {minutes played, goals, assists}. Then, for a member of our data set, the actual Input → Output model would look something like “(6.2, 195, 29, 8) → (22, 0.8, 1.1)”. So, given our set of input nodes and output nodes, the hidden nodes of our NN would have to “figure out” how to transform left side into the right side. And this “figuring out” process is the process of training of NN, during which the weights of the hidden nodes are calculated. This process of training a NN is done on the training sub-set of the sample, and once a NN is constructed, it would be validated on the testing sub-set to check the applicability of the model to a larger set. So, after we trained and tested a NN, we get the model of the type “inputs → transformation → outputs”, where, again, “transformation” component is meaningful to the model, but meaningless to an investigator. However, once it was generated, it could be used for two (and here we concentrate on the most obvious ones) interesting purposes. First, we could predict the outputs of new entities that were not a part of the original set. If we consider a new NHL player, then we could use the developed NN model to predict, based on the original conditions, his performance in terms of the predefined outputs. We would not be able to say how the results (outputs) are obtained, or what the player should do to improve the outputs, but we will be able to come up with the outputs based on the NN model’s “black box” of hidden nodes. The second scenario involves building two NN models and requires a little introduction. Let us consider two groups of NHL players – “flops” and “stars”. “Flops” play less, and “flops” score and assist less. “Stars” play more, and “stars” score and assist more. If we are to consider the situation of “flops”, then there are two plausible explanations for their inferior performance. First, they score and assist less because they play less. So, if we want to improve their performance, we have to let them play more – as much time as the “stars” get to play. Then we’ll have a level playing field for a comparison. But, second, it is also possible that “flops” score and assist less because they are simply inferior players. And before giving “flops” more playing time, they have to improve their level of skills. What is the proper way to go about improving “flops” performance? If we have two NNs generated – one for “stars” and another for “flops”, then we could use it to gain sufficient insight to select a better way. First, we can save “black box” model of “stars” and apply it to the inputs of “flops”. This is analogous to giving “flops” skills of “stars”, and this will allow us in determining whether the poor performance of “flops” is due to their lack of skills. Second, we can apply the inputs of “stars” to the “black box” model of “flops”. This is analogous to giving “flops” inputs of “stars” (e.g., giving more playing time), and this will allow us in finding out if the inferior performance of “flops” is due to their relatively smaller playing time. By comparing the values of the output nodes for both scenarios, we can determine the reason for the difference in the outputs and design an appropriate intervention accordingly.

14

Quantitative Methodologies using Multi-Methods

ASSOCIATION RULES MINING (ARM) There is an interesting similarity between the ARM technique and NNs. Conceptually, both NN and ARM work with, or are comprised of, three fundamental components – left-side variables, right-side variables, and a “black box” in between. Unlike NN, however, which aims to connect specified by the investigator inputs of the left side with outputs of the right side, the goal of ARM is to discover possible associations between left-side variables and right-side variables that are not specified a priori. Thus, while the purpose of the “black box” of a NN is to force the connection between the pre-defined inputs and outputs, the purpose of the “black box” of ARM is to discover associations between the left-side variables and the right-side variables that are not selected in advance. In both cases, the “black box” offers no explanation regarding the nature of the relationship between the left and the right side, leaving this task under the purview of the investigator. Our aim here is in applying ARM for the purposes of Market Basket Analysis (MBA), which gives us an opportunity to discover a set of interesting “If → Then” rules. We must note that “→” does not connote a causal relationship between the left side and the right side, but simply indicate that “If” and “Then” go together. For the purposes of MBA, we have to present our data set as a collection of transactions, where each transaction consists of a set of items. Commonly, a data set for MBA comes from the transaction-level data associated with purchases – “market baskets”. Consequently, a data set for MBA is comprised of the records that are “baskets” of purchases, and we’ll have as many baskets as there are records in the data set. In simple terms, the purpose of MBA is to find out what items in the customer’s basket tend to go together. Clearly, some items, while not necessarily are random purchases (e.g., those that are a result of spur of the moment decision), are caused to be purchased by random events. For example, a person shopping in a grocery store may get a pack of light bulbs (because prior to going to the store, he accidentally tipped over a table lamp and the light bulb broke) or a jug of a windshield washing liquid (because the warning sign came up in her/his car but the local auto shop is already closed). Such purchases are not very interesting for decision makers, because it is hard to predict them, as it is hard to predict random events. However, some purchases do tend to go together – milk and bananas, salad mix and salad dressing, bread and cheese, as well as many other combinations of products. Clearly, in no way or form a purchase of one item causes, deterministically, the purchase of other items, but, nevertheless, they seem to go together naturally. Discovery of such rules is the main goal of MBA, for a decision maker is provided with the actionable information that is based on non-random events, and, therefore, is somewhat predictable. ARM and MBA do not work with numeric data; however, a data set comprising the numeric values could be converted into a nominal data set that could be subjected to the analysis. We will demonstrate to our reader, later in the text, one of the ways to perform such conversion. One of the most popular algorithms for generating “if → then” rules (commonly referred to as itemset) is A priori, and commonly used practice of generating itemsets is based on the threshold values of three criteria – support, confidence, and lift.

Components of Multi-Method Methodologies

15

Support of an itemset is an indicator of a frequency of occurrence of a given rule within the data set. If a value of support equals to 1, then the itemset appears in 100% of the records (transactions) in our data set. Clearly, a value of support could not be equal to 0, but higher the value within the range between 0 and 1, the greater the percentage of the records that contain the rule. A commonly chosen threshold value for support is 0.2, which means that the algorithm will generate itemsets that appear in at least 20% of the transactions (records) in the data set. Once an itemset has been discovered by the algorithm, it needs to be tested – after all, the rule may, or may not be true. The test of how often the rule has been shown to be true within the data set is reflected by the criterion of Confidence. In the case of the value of confidence being 1, this means that all (100%) the transactions (records) containing “If” part also contain “Then” component. Similarly, if the confidence is equal to 0.1, then only 10% of the transaction containing “If” also contain “Then”. Support and Confidence are useful measures for evaluating a likelihood of occurrences of “If” and “Then” within the data set, but they are not good indicators in regard to “→”. This is where Lift comes to play, being a criterion estimating a probability that “If” leads to, or, serves as an antecedent to “Then”. If the value of Lift is equal to 1, then it means that “If” and “Then” are independent of each other; if the value is greater than 1, then it means that the presence of “If” results, actually, in the presence of less of “Then” and vice versa. Consequently, if the value of lift is less than or equal to 1, then the “If → Then” rule is not very useful to a decision maker. However, if the value of Lift is greater than 1 (the upper range is infinity), then the rule is important to the decision maker because, for all intents and purposes, there is evidence that the presence of “If” is positively associated with the presence of “Then”. Simply put, the criterion of lift serves as an indicator of the importance of the discovered rule for the purposes of prediction. Let us consider our running example of the set of NHL players and use of MBA to gain possible insights into the data. In order to generate association rules, we need to have a data set consisting of transactions; thus, we need to convert all the numeric stats into some sort of categories. For example, in regard to playing time, we could generate such categories as “low”, “medium”, and “high”. Or, we could create categories based on four quartiles – “low”, “mid-low”, “mid-high”, and “high”. This process of finding a suitable categorization is under the purview of the investigator; thus, no “one size fits all” scheme exists. Also, in order to generate a set of rules, we must have a much richer data set than the one of the type “minutes played → points”. Instead, we need a data set where we would have a wide variety of attributes for each player. If we consider “minutes played → points” being an analytical model describing transformation of inputs into outputs for each player, then we should also consider that every player exists/functions in their own context – in their own environment. After all, the same “minutes played → points” model does not discern between offensive and defensive players. And the environment always plays an important role by impacting the level of performance of an open system. Consequently, we would like to have a data set that describes the context within which NHL players exist. This could be any data regarding the team where they play, the type of family environment they have,

16

Quantitative Methodologies using Multi-Methods

anything regarding their hobbies, pets, and preferences. We would definitely like to include any data describing the types of exercises they do, the types of diets they may have, and so on. Once we have such a data set, we should be able to generate rules that may inform us, for example, about interesting associations that characterize “low” time hockey players. Similarly, we should be able to generate rules that associate some of the common contextual variables for “star”-level players. As a result, we could obtain some insights into contextual commonalities that the players of different types possess. Such rules, of course, would not allow for explaining how the “playing time” is converted into “points”, but they will offer insights regarding what environmental variables are worth paying attention to.

DATA ENVELOPMENT ANALYSIS (DEA) DEA is a very popular method commonly employed for the purpose of calculating levels of relative efficiency of decision-making units (DMUs). As long as an entity could be described as a set of inputs and outputs, it qualifies as a DMU and could be subjected to DEA. The DMUs in the set do not have to be of the same size or scale, nor do they have to share a similar context. Instead, DMUs have to be semantically similar – they have to be of the same type of systems that receive inputs and produce outputs. For example, it would be reasonable to compare, from DEA perspective, different types of restaurants, airports, and factories, but it would be hard to defend a comparison of restaurants vis-à-vis airports vis-à-vis factories. Simply put, DMUs in the sample have to make sense to be grouped together for a meaningful comparison. So, if we consider a group of semantically similar entities (e.g., hockey players, departments, firms, and hospitals) that are described by the same DEA model – by the same set of inputs and outputs, then we can subject that data set to DEA to compare their levels of relative efficiency. There are few points worth mentioning regarding a DEA model. First, inputs and outputs of DEA are non-monetary, for it is not the purpose of DEA to accommodate a model of the type “investments → revenues”. Instead, inputs and outputs are those other variables than costs and prices that make sense to the investigator. For example, we can compare efficiencies of a group of students using the following DEA model: “(#ofStudyDays, #ofHoursPerStudyDay) → TestGrade”. Similarly, we would not use DEA to conduct an analysis of relative efficiencies of a group of bakeries using a model “cost of ingredients of a cake → price of a cake”, but we could use DEA to investigate a model “(# of employees, # of shift hours) → # of pastries baked”. Second, the inputs and outputs are not necessarily representative of a productiontype “Input → Output” model. For example, we would not be interested in a DEA model consisting of “weight of ingredients to make a cake → weight of a cake” to evaluate the level of a relative performance of a set of bakeries, but we may use a DEA model consisting of “(# of employees, # of hours of operation) → # of cakes sold”. Once DEA modeling is adopted, all DMUs are considered to be completely described by the inputs and outputs of the model – no context-specific, environmentally relevant attributes are considered by the analysis.

Components of Multi-Method Methodologies

17

Fundamentally, the approach of DEA is that of constructing a ratio – the multiple inputs and outputs of DEA model are collapsed/transformed into one meta-input and one meta-output, and the ratio of the two is expressed as a score of a relative efficiency. It is important to note that DEA does not compare each DMU with the rest of the DMUs in the sample, but only with a suitable sub-set comprised of its peers. The DMUs with the highest values for the ratio are considered to be relatively efficient as compared to relatively inefficient DMUs in the set. All relatively efficient DMUs receive a perfect score of “1”, which indicates a 100% efficiency. Once a sub-set of relatively efficient DMUs is identified, they form an efficiency frontier that envelops the rest of the relatively inefficient DMUs. Another aspect of DEA that is worth mentioning is that of orientation of the model. DEA allows for three orientations: Input-Oriented, Output-Oriented, and Base-Oriented. Input-oriented model is used in the scenarios where the inputs are controllable (e.g., students can control how many days a week and how many hours per day they study), Output-Oriented model is suitable for situations when the outputs are controllable (e.g., bakery can control how many cakes it is going to produce), and Base-Oriented model assumes the control over the inputs as well as the outputs (e.g., a writer has a control over how many hours a day she is going to dedicate to writing, as well as she can control the number of pages written). Regardless of orientation, all relatively efficient DMUs receive scores of 1, but relatively inefficient DMUs will receive scores of less than 1 in the case of input orientation, and greater than 1 in the case of output orientation. Additionally, there are options regarding return-to-scale, where DEA offers to an investigator consideration for constant, variable, and non-increasing (decreasing) returns to scale. The choice of an option impacts the shape of the efficiency frontier, so, for example, a relatively efficient DMU under constant return to scale is also efficient under variable return to scale, but the opposite is not necessarily true. DEA is a point-in-time method, where calculations of the scores of relative efficiencies of DMUs are undertaken for a chosen temporal snapshot. However, it is also possible to use DEA to investigate changes in the scores of relative efficiencies of the DMUs over time. This is done via calculating values of Malmquist Index (MI), which measures the changes in scores that took place over a period of time – let us say, year 1 and year 2. Fundamentally, values of MI are calculated by conducting DEA twice – once for year 1 and the second time for year 2. MI is a very useful tool, for it allows an investigator to assess whether, over a period of time, a given DMU became more efficient (positive change in productivity), stayed the same (no change), or became less efficient (drop in productivity). This, on its own, is very valuable, for we can assess (let us continue using our example of NHL players) not only relative efficiency of the player in regard to conversion of minutes played into points but also whether a hockey player became better, stayed the same, or became worse over a period of time. There is a great additional benefit offered by MI to an investigator – it allows to identify the sources of the change that took place over time. Let us consider the approach of DEA – the approach of constructing an efficient frontier at a given point in time. For all intents and purposes, DEA tells us how far away from the efficient frontier a given DMU is situated. But the approach of MI is based on measuring the movement of the DMUs over time, and there are two components that comprise the movement.

18

Quantitative Methodologies using Multi-Methods

First, there is a movement of the efficiency frontier itself – this movement is represented by Technology Change (TC) component of MI. Second, there is a movement of a DMU relative to the frontier – this is Efficiency Change (EC) component of MI. If we, again, consider our example of NHL players, then MI would be able to provide us some very valuable insights. Let us say, that a given player exhibited a positive change in his score of relative efficiency over the period of one year – his value of MI is greater than 1. What caused the change? Well, it is possible that the positive change was a result of the movement of the frontier – the player was “dragged” up by the technological advancements that took place over time. The skates became better, the stick became lighter, a better helmet, and so on. But it is also possible that the positive change was a result of the improved skills of the player – he improved his skating, he improved his puck handling, and so on. By calculating MI and its components, we would be able to find the reason for the improvement. Similarly, if a given player decreased its score of the relative efficiency, then we’ll be able to identify the sources responsible for the deterioration. Clearly, MI is a useful tool used for the purposes of designing an intervention directed toward improvement of the level of relative efficiency of DMU, because it offers a justification for the targeted allocation of resources. For example, if a given hockey player exhibited a decrease in the level of the relative efficiency over time (e.g., MI1) and the negative change in efficiency (e.g., EC 1 for the given economy then it has exhibited growth in productivity. If the corresponding EC > TC then this growth in productivity is due to improvement in Efficiency; while if the corresponding TC > EC then this growth in productivity is due to improvement in the application of the Technology.

133

Micro-Economic Impacts of ICT Capabilities

Phase 2: Decision Tree-Based Analysis The second step of the methodology involves DT analysis. The purpose of the DT induction is to identify the input variable that is selected for the top-level partitioning of the data set. Specifically, DT analysis will allow for answering the following questions described below in Steps 1 and 2 of this phase: Phase 2, Step 1 1. What are the differences between the groups of economies in our sample in terms of ICT-related factors and outcomes for each year? To answer this question, we use DT induction to generate separate DTs of depth 2 (i.e., each rule will have at most two predictor variables) for the years 2010 and 2011. It should be noted that our input data set contains an Economy Group variable that indicates the Economy Group associated with the given row of the data set. In this application of DT induction, the Economy Group variable is used as the target variable, and the component variables associated with ICT Capabilities (see Table 20.2) are used as the potential predictor variables. The reader may recall the DT induction process involves recursive partitioning of the data set based, where the split of the potential predictor variable that provides the top value of the relevant splitting measure is used for the first partitioning of the data set into its first-level sub-sets, and that subsequent partitioning of these sub-set involves the same approach. The resulting DT will thus have branches that are associated with the splits of the given variable, and nodes that provide the associated relative frequencies for the Economy Groups. Our aim in using DT induction is to identify the nodes of the DT in which one of the Economy Groups has a relative frequency of at least 50%. This will then allow us to identify the range of values of the top ICT Capabilities component variables that are associated with each Economy Group. It should, however, be noted that based on the actual data it is possible that for a given Economy Group there may not be any node of the resulting DT for which it has a relative frequency of at least 50%. This would indicate that sufficiently strong differentiating factors were not identified.

TABLE 20.2 Groups of Countries (Based on IMF Classification of 2011) and Membership of Each Group Advanced Economies Estonia Slovenia Czech Republic Slovak Republic Spain

Central and Eastern Europe Hungary Latvia Poland Lithuania Montenegro

Commonwealth of Independent States and Mongolia Armenia Kazakhstan Moldova Kyrgyz Republic Tajikistan

Middle East and North Africa Morocco Tunisia Algeria Oman

Sub-Saharan Africa Kenya Ghana Senegal Namibia Nigeria

134

Quantitative Methodologies using Multi-Methods

Phase 2, Step 2 2. What are some of the differentiating characteristics of the Relatively Efficient and Inefficient economies? To answer this question, we use DT induction to generate separate DTs of depth 2 (i.e., each rule will have at most two predictor variables) for each of our DEA models. It should be noted that our input data set is extended to contain an Efficiency_Class variable that indicates whether the Economy Group associated with the given row of the data set is relatively Efficient or relative Inefficient. In this application of DT induction in this phase, the Efficiency_Class variable is used as the target variable, and the component variables associated with ICT Capabilities and Microeconomic Outcomes (see Table 20.2) are used as the potential predictor variables. Here our aim in using DT induction is to identify the nodes of the DT in which one of the Efficiency_Classes has a relative frequency of at least 50%. This will then allow us to identify for each of our DEA Models, the range of values of the top ICT Capabilities and Microeconomic Outcomes component variables that are associated with each Efficiency_Class. It should, however, be noted that based on the actual data it is possible that for a given Efficiency_Class there may be not be any node of the resulting DT for which it has a relative frequency of at least 50% which would indicate that sufficiently strong differentiating factors based were not identified.

DESCRIPTION OF THE DATA We obtained the data from two sources – the database of the World Development Indicators (WDI) and Global Information Technology Reports of 2010 and 2011. A twoyear period, while may not being sufficient for the purposes of performing an in-depth time series data analysis, is sufficient for the purposes of illustrating and evaluating the proposed framework and associated methodology. Overall, we compiled the data on 24 economies of the world representing five groups according to the classification of the International Monetary Fund 2011 (Nielsen, 2011). Membership of each group is provided in Table 20.2. The NRI framework does not split the set of the countries into the various sub-groups when assigning scores to the subindexes. Consequently, our partitioning of the sample of 24 countries into 5 groups is logical in nature. The limited number of the economies of this study is due to exploratory nature of the investigation – once the framework is developed and tested we will increase the sample size and conduct the follow-up study in a large context.

RESULTS OF THE DATA ANALYSIS Results from Phase 1: Application of Data Envelopment Analysis (DEA) Phase 1, Step 1 We start by addressing the second research question RQ2: What areas of ICT may require innovative applications of the available resources for each group?

135

Micro-Economic Impacts of ICT Capabilities

TABLE 20.3 Average Relative Efficiency Score for Each Model Model Year 2010 2011

I1.O1 0.60 0.55

I1.O2 0.66 0.66

I1.O3 0.62 0.58

I2.O1 0.61 0.60

I2.O2 0.67 0.63

I2.O3 0.63 0.63

I3.O1 0.80 0.78

I3.O2 0.83 0.87

I3.O3 0.84 0.82

In order to do so, for each model we first calculate Relative Efficiency score for each economy for each year, which then allows us to identify the model with the lowest average Relative Efficiency score (see Table 20.3). In our case, for both years, it is a model I1.O1, and the corresponding ICT Capability is Environment (label I1). The results allow us to answer RQ2 as follows: Market Environment, Infrastructure Environment, and Political and Regulatory Environment are the areas of ICT Capabilities that may require innovative application of the available resources for each group. Phase 1, Step 2 We next addressed the question: Which ICT Capability/Microeconomic Outcome path would benefit the most from the changes to the input–output transformation process? In order to do so, we first generated the Malmquist Index (MI) scores for each economy for each model (for details see Table 4A of the Appendix), then calculated the average MI score for each model (see Table 20.4), following which we assessed which of these scores was less than or equal to 1. Given our results displayed in Table 20.4, the answer to the second question for the period of 2010–2011 is: The ICT Capabilities to Financial Well-Being paths would benefit the most from the changes in the Input–Output Transformation Process. An interesting pair of related issues are: • What is the economy group that should be used as a benchmark of efficiency for a given model? • What is the economy group that shows the greatest improvement for a given model? TABLE 20.4 Average Malmquist Index (MI) Score for Each Model Model Period 2010–2011

I1.O1 1.11

I1.O2 0.97

I1.O3 1.03

I2.O1 1.10

I2.O2 0.95

I2.O3 1.02

I3.O1 1.10

I3.O2 1.00

I3.O3 1.01

136

Quantitative Methodologies using Multi-Methods

TABLE 20.5 Benchmark Group for Each Model Benchmark Economy Group Model I1.O1: Environment → Trade I1.O2: Environment → Financial Well-Being I1.O3: Environment → State of Labor Market I2.O1: Usage → Trade I2.O2: Usage → Financial Well-Being I2.O3: Usage → State of Labor Market I3.O1: Readiness → Trade I3.O2: Readiness → Financial Well-Being I3.O3: Readiness → Financial Well-Being

2010 Advanced Economies (AE) AE AE Comm. of Ind. States (CIS) CIS CIS AE AE AE

2011 AE AE AE CIS CIS CIS AE AE AE

In order to make this determination of the benchmark group for efficiency we calculate the average relative efficiency score of each economy group, per model per year. For each model, the economy group that provides the highest average relative efficiency score is then selected as the benchmark group for that year. Table 20.5 identifies the benchmark economy group for each model for each year. Phase 1, Step 3 With regards to our third research question RQ3 (i.e., What are the areas of strength and weakness in terms of the efficiency of the “ICT Capabilities → Microeconomic Outcomes” paths for each group), evaluation of the strengths and weaknesses can be performed using two criteria: (1) whether a given economy exhibited growth in productivity for a given path? (e.g., is MI > 1 or not?); and (2) what is the dominant source of growth in productivity? (e.g., which component is greater in value, EC or TC?). The information presented in Tables 20.6 and 20.7 could be used to answer these questions.

TABLE 20.6 Growth (+)/Decline (−) in Productivity per Economy Group per Model Economy Group AE CEE CIS MENA SSA

I1.O1 + + + + +

I1.O2 − − − − −

I1.O3 + + + + −

I2.O1 − + + + +

I2.O2 − − + + −

I2.O3 − − + + −

I3.O1 + + + + +

I3.O2 − + − + −

I3.O3 − + + + +

137

Micro-Economic Impacts of ICT Capabilities

TABLE 20.7 Economy Group with Greatest Change in Productivity per Model Model I1.O1: Environment → Trade I1.O2: Environment → Financial Well-Being I1.O3: Environment → State of Labor Market I2.O1: Usage → Trade I2.O2: Usage → Financial Well-Being I2.O3: Usage → State of Labor Market I3.O1: Readiness → Trade I3.O2: Readiness → Financial Well-Being I3.O3: Readiness → Financial Well-Being

Economy Group Advanced Economies Middle East and North Africa (MENA) Central and Eastern Europe (CEE) Comm. of Ind. States (CIS) CIS CIS CEE CEE MENA

Dominant Source TC TC TC EC TC EC TC EC EC

Information summarized in Table 20.6 allows for answering RQ2 as follows: With regard to the efficiency of the “ICT capabilities → Microeconomic Outcomes” paths, for each group, the areas of strength are associated with the growth in productivity and the areas of weakness with the decline in productivity. This allows us to conclude that the proposed framework successfully passed the test of the capability to identify those mechanisms of transformation of ICT capabilities into macroeconomic outcomes that need to be adjusted.

Results from Phase 2 – Decision Tree (DT) Based Analysis We now focus on the first research question RQ1, namely: What are the specific characteristics of each group with regards to ICT and the impacts of ICT? To address this question we executed Phase 2. Step 1 (see “Phase 2: Decision Tree Based Analysis” section); the corresponding results are displayed in Table 20.8. The reader may note that differentiating factors were identified for all our economy groups with the with the exception of the Middle East and North Africa (MENA) group. Given the information displayed in Table 20.9 the answer to RQ1 is as follows: The difference between the groups of economies in our sample could be expressed in terms of the differences with regards to the following ICT Capabilities • Market Environment and Individual Usage in 2010, and • Market Environment and Infrastructure Environment in 2011.

138

Quantitative Methodologies using Multi-Methods

TABLE 20.8 Differences between the Groups of Economies in Terms of ICT Capabilities Year 2010

2011

Group Advanced Economies (AE) Central and Eastern Europe (CEE) Comm. of Ind. States (CIS) Middle East and North Africa (MENA) Sub-Saharan Africa (SSA) AE CEE CIS MENA SSA

Differentiating Factors MarketEnv ≥ 4.355 & IndUse ≥ 3.5 MarketEnv < 4.355 & IndUse ≥ 3.5 MarketEnv < 3.925 & IndUse < 3.5 N/A MarketEnv ≥ 3.925 & IndUse < 3.5 InfraEnv > 3.65 InfraEnv > 3.65 MarketEnv < 3.85& InfraEnv < 3.65 N/A MarketEnv ≥ 3.85& InfraEnv < 3.65

Classification 100% of AE 100% CEE 83% of CIS N/A 63% of SSA 100% of AE 100% of CEE 83% of CIS N/A 63% of SSA

It can be seen that the information that results from the execution of Phase 2. Step 1 could be used by policymakers/decision-makers for formulating improvement strategies. For example, if an economy-member of the SSA group aims to benchmark a performance of an economy-member of the Advanced Economies (AE) group, then the results show that in regard to ICT Capabilities it could be wise to direct attention to Individual Usage, for this is the area where two groups of economies are clearly different. This may mean increasing the level of ICT investments in this area, or applying existing levels of investments in innovative way. We now focus on the second research question of this phase: What are some of the differentiating characteristics of the relatively efficient and relatively inefficient economies?

TABLE 20.9 DT Analysis: Characteristics of Efficient and Inefficient Economies, per Model Model Environment → Trade Environment → Income Environment → Labor Readiness → Trade Readiness → Income Readiness → Labor Usage → Trade Usage → Income Usage → Labor

Characteristics of Efficient Economies N/A

Characteristics of Inefficient Economies N/A

BusRead ≥ 4.245 Exports ≥ 43.4345 IndRead ≥ 4.51 GovRead ≥ 4.345 IndRead ≥ 4.545 & BusRead y” for high school players, “if (experience > n) then (assists > m)” and so on. So, the trick is to identify a set of rules that is based on a set of common criteria that differentiates the groups – this insight is provided by DTI. Consequently, the novelty of our approach is associated with its capability to identify the main differentiating factors responsible for heterogeneity of the context, and then to base the selection of the rules on those factors. Previous investigations used a hybrid DEA/DTI methodology (Samoilenko & Osei-Bryson, 2007), and the use of ARM with DEA was recently reported by Samoilenko (2016); however, this investigation represents the first case of using the three methods (i.e., DEA, DTI, ARM) in synergy. Simply put, if DEA allows us to identify the efficient performers, and DTI helps us to discover the relevant dimensions that differentiate efficient and inefficient performers, then ARM allows us to benchmark efficient performers via a set of “IF THEN” rules that rely on the discovered by DTI dimensions. To our knowledge, no other combination of data analytic and data mining methods could offer so much in so few steps.

147

Socio-Economic Outcomes of ICT Capabilities

Phase 1: Data Envelopment Analysis (DEA) During the first phase we rely on DEA to evaluate relative efficiency of three “Drivers → Impact” paths. We will use variable return to scale (VRS) DEA model to conduct the analysis, for it is reasonable to argue that SSA economies have not yet reached the point of developing a level of ICT infrastructure allowing accruing the benefits yielded by capitalizing on economies of scale. Given a four-year time period we will run DEA 12 times. Consequently, for each economy in the sample we are going to have four scores of relative efficiency for each of the three models. At this point we need to provide a justification for the inputs and outputs included in our models. In regard to outputs the reasoning is intuitive – first, we would like to assess the efficiency of the overall impact, and then, each type of the impact separately. This is because an economy could be efficient in obtaining one type of an impact (e.g., economic) and not efficient in regard to another impact (e.g., social) (Table 21.1). With regard to the choice of the inputs of DEA model, our approach is methodological. While we are free to use eight sub-categories of Drivers as inputs of a DEA model, the general rule of thumb is that for a reasonable level of discrimination number of economies (or Decision Making Units in DEA terms) must be at least twice the product of inputs and outputs (Dyson, Allen, Camanho, Podinovski, Sarrico, & Shale, 2001). In our case we have a sufficient number of economies in our set, but if we use a DEA model with eight inputs and two outputs then we would need to have at least 2 × 8 × 2 = 32 economies in the sample. Furthermore, and more importantly, the greater the number of factors included in the DEA model, the lower the level of discrimination of the model (Dyson et al., 2001). However, we would like to use all the data available to us so we could inquire, for example, whether a set of specific factors – pillars – differentiate relatively efficient economies from relatively inefficient once. We would use DTI to do so. Additionally, we use DEA to calculate the values of the Malmquist index (MI) – this allows us to assess the changes in relative efficiency of SSA economies that took place over time. Such results would not only identify the economies that exhibited growth

TABLE 21.1 DEA Models of the Study DEA Model Drivers → Overall_Impact (DOI) Drivers → Economic_Impact (DEI) Drivers → Social_Impact (DSI)

Inputs of DEA Model Environment Subindex Readiness Subindex Usage Subindex Environment Subindex Readiness Subindex Usage Subindex Environment Subindex Readiness Subindex Usage Subindex

Outputs of DEA Model Impact Sub-index

Economic Impact Sub-category

Social Impact Sub-category

148

Quantitative Methodologies using Multi-Methods

in productivity (under assumption of constant return to scale), but to also identify the sources of growth (EC – change in efficiency vs. TC – change in technology). By applying DTI, we can identify factors that differentiate the Growth vs. No Growth economies.

Phase 2: Decision Tree Induction (DTI) To proceed with Phase 2 we need to create a new variable “Target” to differentiate various groups of economies. We are interested in three types of groupings: first, we would like to differentiate the groups of SSA by their level of economic development, and then we would like to differentiate the groups in terms of their efficiency of the socio-economic impact of ICT. The last analysis would involve differentiating SSA economies by growth in productivity – Growth vs. No Growth. Thus, we would conduct DTI three times, which would require Target to have three domains of values. In the first case, grouping by income, the domain of values of Target would be {1, 2, 3}, for, respectively, Low Income (LI), Low Middle Income (LM), and Upper Middle (UM) groups of economies. In the second case, grouping by efficiency, Target would assume the values of {0, 1}, for, respectively, relatively inefficient, and relatively efficient SSA. The same domain of values, namely, {0, 1}, could be applied to the grouping by growth in productivity, where “0” would indicate “No Growth” and “1” would indicate “Growth”.

Phase 3: Association Rule Mining (ARM) The purpose of Phase 3 is to find possible patterns, associations, or causal structures that may exist in our data. One of the main advantages of ARM is that it is suitable for undirected data mining; thus, we’ll aim to discover naturally occurring associations between the factors (sub-indexes of Drivers and Impact) – components of NRI. ARM could be classified as either being explanatory or exploratory in nature. In the case of our investigation we employ exploratory ARM, for we do not have any theoretical support for why certain relationships between the sub-indexes of NRI should exist. A very common approach to generating associations between the variables, or itemsets, via ARM is by using the a priori algorithm (Agrawal & Ramakrishnan, 1994) – we will rely on this approach in the current investigation. Transformation of the data is required for this step – we follow the method of Samoilenko (2016) to do so.

RESEARCH QUESTIONS AND NULL HYPOTHESES OF THE STUDY At this point, we can operationalize the two objectives of this investigation in the form of the specific research questions and corresponding null hypotheses. The first research question operationalizes the first objective as follows: Is the developed methodology capable of generating sets of differentiating factors and association rules for a given set of criteria? One of these criteria is associated with the level of income of economies (e.g., Low Income vs. Low-Middle vs. Upper-Middle), while another criterion is a relative level

Socio-Economic Outcomes of ICT Capabilities

149

of efficiency (e.g., relatively efficient vs. relatively inefficient) of Drivers → Impact path, and the third one is a growth in productivity that took place over period of time. We can answer this research question by testing the corresponding null hypotheses: H01a: The DTI part of the methodology will fail to generate a set of differentiating factors characterized by high-level splits. We will test H01a under the conditions of: high-level splits that differentiate at least 60% of at least one of the groups of SSA economies. H01b: The ARM part of the methodology will fail to generate sets of association rules for a given set of criteria. We will test H01b under the minimal conditions of Support > 20%, Confidence > 1.0, and Lift > 1.0. While the results of DTI and ARM may offer useful insights by themselves, we would like to use the two methods in a complementary fashion; thus, we state another hypothesis as follows: H01c: The results of DTI and ARM are not complementary. We will test H01c under the condition that the differentiating factors identified by DTI would be included in sets of rules identified by ARM. The second research question operationalizes the second objective as follows: Does the choice of criteria such as level of economic development, relative efficiency, and growth in productivity impact the combination of factors describing various groups of economies and relationships between Drivers and Impacts of ICT? Basically, we would like to find out if the different criteria could be characterized via different set of factors – this allows us to inquire into the specificity of a setting expressed as a combination of sub-categories of NRI. We will answer the second research question after testing our second null hypothesis: H02: No combination of factors contained in the generated association rules would be unique to a given context. The simple side-by-side comparison of the generated association rules and split variables will serve as a sufficient criterion for testing H02.

THE DATA We obtained the data from a reputable source – the World Economic Forum’s Global Information Technology Report 2015 (GITR, 2015). In 2012 the representation of NRI was partially changed in terms of the number and representation of the pillars

150

Quantitative Methodologies using Multi-Methods

TABLE 21.2 Sample of Sub-Saharan Economies, by Income Level Income Level Low Income Low-Middle Income Upper-Middle Income

Sub-Saharan Economies Burkina Faso, Burundi, Chad, Ethiopia, Gambia, Kenya, Madagascar, Malawi, Mali, Mozambique, Rwanda, Tanzania, Uganda, Zimbabwe Cameroon, Cape Verde, Côte d’Ivoire, Ghana, Lesotho, Nigeria, Senegal, Swaziland, Zambia Botswana, Mauritius, Namibia, South Africa

of three sub-indexes of NRI; it was also the year when the Impact subindex was introduced. Given the changes that took place between 2011 and 2012, we decided to concentrate on the new version of NRI and collect the data provided in GITR 2012, 2013, 2014, and 2015. In some cases, the representation of SSA economies was inconsistent – for example, we could not include Angola, Seychelles, Liberia, Gabon, Sierra Leone, and Guinea in our sample because the data for some of the years was missing. While there is an advantage to increasing the sample size of a study, there is a price to pay via dealing with missing variables, imputation of values, and additional data preprocessing. After considering the pluses and minuses of “sample size vs. data actually available” we have assembled a smaller data set that contained no missing data and no outliers, but was as reliable as one could get from a given source. Overall, we were able to compile the data set representing 27 economies of SSA (the classification of the International Monetary Fund as of October 2014). The sample consists of 14 low income economies, nine low-middle economies, and four upper-middle economies (the classification of the World Bank as of July 2014). Membership of each group of the sample is provided in Table 21.2.

RESULTS OF THE DATA ANALYSIS Phase 1: Data Envelopment Analysis We offer a summary of the results of DEA below. If a given economy has been determined to be relatively efficient for at least three times over the period of four years, we have labeled such economy as “efficient” for the whole period of four years. Because our economies fall within three distinct groups – low income (LI), low-middle income (LM), and upper-middle income (UM), we also determined the relative efficiency of each economy over the 4 years within its group – we will use this information in Phase 3 when we perform ARM. Our results demonstrated that seven economies out of the full sample are relatively efficient with regard to the impact of Drivers on social, economic, and overall Impact of ICT. Additionally, we identified relatively efficient economies per each of the income-level group; in some cases (e.g., Burundi, Chad, Kenya, Mali, Rwanda, and Senegal), the relatively efficient within its group’ economies are also efficient overall. In other cases (e.g., Swaziland, Lesotho, Botswana, Mauritius, Namibia, and

151

Socio-Economic Outcomes of ICT Capabilities

TABLE 21.3 Results of DEA Income Level LI

LM

UM

Economy BFA BDI TCD ETH GMB KEN MDG MWI MLI MOZ RWA TZA UGA ZWE CMR CPV GHA NGA SEN SWZ ZMB CIV LSO BWA MUS NAM ZAF

Overall Efficiency of the Impact of ICT Efficient Efficient Efficient Efficient Inefficient Efficient Inefficient Inefficient Efficient Inefficient Efficient Inefficient Inefficient Inefficient Inefficient Inefficient Inefficient Inefficient Efficient Inefficient Inefficient Inefficient Inefficient Inefficient Inefficient Inefficient Inefficient

Overall Changes in Productivity Growth Growth No growth No growth Growth No growth Growth No growth Growth No growth No growth Growth No growth Growth Growth No growth No growth No growth No growth Growth Growth Growth Growth No growth Growth No growth No growth

Growth via EC? Yes Yes No No No No Yes No No Yes Yes Yes Yes Yes Yes No No No No Yes Yes Yes Yes No Yes Yes No

Growth via TC? Yes No No No No No No No Yes No No No No No No No No No Yes No No No No No No No No

South African Republic), the relatively inefficient, overall, economies end up being efficient within their respective group (Table 21.3). Additionally, we used DEA to calculate the values of Malmquist Index (MI), which allows us to assess the changes in the scores of relative efficiency that took over period of time. Under the assumption of constant returns to scale the change indicates changes in productivity. Consequently, we are able to assess whether the economies become more productive or not. Because MI is comprised of two components – change in efficiency (EC) and change in technology (TC), we are also able to assess whether the changes in productivity are associated with a particular component. Overall, only 11 economies (40% of the sample) exhibited growth in productivity.

152

Quantitative Methodologies using Multi-Methods

An analysis of the changes in EC and TC offers an interesting insight: 16 economies (60% of the sample) exhibited positive changes in efficiency, but only 3 economies (11% of the sample) demonstrated positive changes in technology. Finally, it is worth noting that only one economy, Burkina Faso, exhibited a balanced growth in productivity, when the growth was driven by both components of MI. Overall, the picture suggests that SSA, as a group, would benefit from a better technology – this suggests that investments in ITC infrastructure should be prioritized. Finally, it is worth noting that eight economies (30% of the sample) have not only exhibited a decline in productivity, but have exhibited a decline in terms of both components of MI. Overall, we summarize the results of DEA part of our methodology as follows: • Seven economies out of the full sample are relatively efficient with regard to the impact of Drivers on social, economic, and overall Impact of ICT • There are relatively efficient economies per each of the income-level group • Only 11 economies (40% of the sample) exhibited growth in productivity, while 8 economies (30% of the sample) have exhibited a decline • SSA, as a group, would benefit from a better technology.

Phase 2: Decision Tree Induction The results of DTI allow us to test our first null hypothesis, H01a, for decision tree induction did generate high-level splits that differentiated groups of economies. Results summarized in Table 21.4 show that such pillars of NRI as Individual Usage, Business Usage, and Skills Readiness do play important role in differentiating three groups of economies. It is not surprising that there is appear to be a clear-cut difference between Low Income and Upper-Middle Income economies, and much less of a difference between Low-Middle Income economies and the other two. We could also identify Individual Usage and Economic Impact as pillars that play role in differentiating relatively efficient SSA economies from inefficient ones. It appears that Infrastructure Readiness, Affordability Readiness, and Individual

TABLE 21.4 Results of Decision Trees Analysis Grouping by Economic development

Group Low Income vs. Low Middle Income vs. Upper Middle Income

Relative efficiency

Relatively Efficient vs. Relatively Inefficient Growth vs. No Growth

Change in productivity

Differentiating/Split Variable Individual Usage Business Usage Skills Readiness Individual Usage Economic Impact Infrastructure Readiness Affordability Readiness Individual Usage Social Impact

153

Socio-Economic Outcomes of ICT Capabilities

Usage are factors playing role in differentiating those economies that became more productive from those that didn’t. At this point, we summarize the results of DTI part of our methodology as follows: • Pillars of NRI such as Individual Usage, Business Usage, and Skills Readiness do play important role in differentiating three groups of economies • Individual Usage and Economic Impact are pillars that play a role in differentiating relatively efficient SSA economies from inefficient ones • Infrastructure Readiness, Affordability Readiness, and Individual Usage differentiate those economies that became more productive from those that didn’t.

Phase 3: Association Rule Mining The results summarized in Table 21.5 allow us to test our null hypotheses. First, the results allow us to reject H01b, for the application of ARM did result in the generation of multiple association rules under the criteria of Support > 20%, Confidence > 1.0, and Lift > 1.0. Second, the results also allow us to reject H02, for the generated by ARM rules contain context-specific combinations of factors. TABLE 21.5 Impact-Specific Rules for Low Income and Low-Middle Income economies Condition LI: Low Income

LM: Low-Middle UM: Upper-Middle

Low Income, Inefficient

Low-Middle, Efficient

Generated Rules low IND_USE, low BUS_USE low SKILL_READ, low BUS_USE midhigh BUS_USE midhigh SKILL_READ high SKIL_READ, high AFFORD_READ, high GOV_USE high BUS&INNOV_ENV, high INFR_READ, high BUS_USE high BUS&INNOV_ENV, high IND_USE, high BUS_USE low INFR_READ, low BUS_USE low IND_USE, low BUS_USE low BUS_USE, low GOV_USE low BUS_USE, low ECON_IMP



low ECON_IMP

Sup. 21%

Conf. 0.7

Lift 1.9



low ECON_IMP

21%

0.6

1.5

→ → →

midlow SOCIO_IMP midhigh SOCIO_IMP high SOCIO_IMP

20% 25% 32%

0.5 0.5 1.0

2.5 1.5 2.3



high ECON_IMP

31%

0.85

2.3



high ECON_IMP

31%

0.85

2.2



low SKILL_READ

36%

1.0

1.9



low SKILL_READ

32%

1.0

1.9



low ECON_IMP

25%

1.0

4.0



low SOCIO_IMP

25%

1.0

3.0

154

Quantitative Methodologies using Multi-Methods

Finally, the results summarized in Table 21.5 allow us to reject H01c that the results of DTI and ARM are not complementary. DTI identified Individual Usage, Business Usage, and Skills Readiness as the variables differentiating the groups of economies in our sample.

DISCUSSION OF THE RESULTS The results of the data analysis, presented in the previous sections, offer evidence that we were successful in addressing the research questions of this study. We developed and tested a novel methodology allowing for investigating complex contextspecific relationships between the factors reflecting the state and the impact of ICT capabilities. The discussion of the results is presented along the points that we considered noteworthy. First, Despite the presence of complex relationships between the Drivers and Impacts of ICT there are common themes associated with the levels of the scores of factors comprising NRI – Business Usage, Individual Usage, and Skills Readiness appear to have a direct relationship with the levels of the scores of socio-economic Impact of ICT. We point out that while the variety of association rules has been generated for a different set of criteria, a common line could also be glanced – some subcategories of NRI’ subindexes (e.g., related to Skills, Business, Individual usage) appear more frequently than other subcategories. Second, Results of our investigation suggest that Business Usage and Individual Usage are among the factors that appear to differentiate economies in terms of their level of economic development, as well as in terms of their relative efficiency of the impact of ICT on the socioeconomic bottom line. These results suggest that wealthier and more efficient economies tend to have higher scores of Business Usage and Individual Usage. The presence of a simple association between the level of income of an economy, its efficiency, and ICT usage seems to be apparent. Third, Infrastructure and Affordability of ICT seem to have an impact on growth in productivity of SSA economies. This finding is important because it was provided by two different methods of analysis – DEA and DTI. According to the results of DEA only three economies exhibited growth in technology over the period 2012–2015, and DTI independently confirmed it by identifying Infrastructure and Affordability as factors differentiating growth vs. no growth economies of SSA.

Socio-Economic Outcomes of ICT Capabilities

155

CONCLUSION In this investigation, we developed and applied a methodology allowing for generating sets of association rules from the combination of factors describing relationships between Drivers and Impact of ICT. The results of the data analysis do confirm the notion that the relationships between the factors representing Drivers and Impact are indeed complex. However, the underlying complexity of the relationship could be made more transparent to researchers and practitioners by the developed in this study methodology.

ACKNOWLEDGMENT Material in this chapter previously appeared in: A methodology for identifying sources of disparities in the socio-economic impacts of ICT capabilities in SubSaharan economies. In International Conference on Information Resources Management (CONF-IRM). Association for Information Systems.

REFERENCES Agrawal, R., and Ramakrishnan, S. (1994). Fast Algorithms for Mining Association Rules in Large Databases. In Proceedings of the 20th International Conference on Very Large Data Bases (VLDB ‘94), Jorge B. Bocca, Matthias Jarke, and Carlo Zaniolo (Eds.). Morgan Kaufmann Publishers Inc., San Francisco, CA, pp. 487–499. Di Battista, A., Dutta, S., Geiger, T., and Lanvin, B. (2015). The Networked Readiness Index 2015: Taking the Pulse of the ICT Revolution, Global Information Technology Report 2015, pp. 3–28. Dutta, S., and Jain, A. (2003). The Networked Readiness of Nations. In The Global Information Technology Report 2002–2003, Dutta, S., A. Lanvin, B., & Paua, F. (eds.). Oxford University Press, New York, NY, Oxford. Dutta, S., Geiger, T., and Lanvin, B. (Eds.) (2015). Global Information Technology Report 2015. ICTs for Inclusive Growth. World Economic Forum and INSEAD, Geneva. Retrieved April 20, 2015 from: http://www3.weforum.org/docs/WEF_Global_IT_ Report_2015.pdf. Dyson, R.G., Allen R., Camanho A.S., Podinovski, V.V., Sarrico C.S., and Shale, E.A. (2001). Pitfalls and Protocols in DEA. European Journal of Operational Research, 132, 245–259. Eide, E.B. (2015). Preface. In Global Information Technology Report 2015, p. v. GITR. (2015). World Economic Forum’ Global Information Technology Report. Available on-line at: http://reports.weforum.org/global-information-technology-report-2015/ network-readiness-index/. Samoilenko, S. (2016). Disparity of Social and Economic Impact of ICT Capabilities in SubSaharan Economies: Empirical Investigation of Differentiating Factors, in Proceedings of the SIG GlobDev Pre-ECIS Workshop ICT in Global Development; Istanbul, Turkey, June 12, 2016. Samoilenko, S., and Osei-Bryson, K.M. (2007). Increasing the Discriminatory Power of DEA in the Presence of the Sample Heterogeneity with Cluster Analysis and Decision Trees, Expert Systems with Applications, 34, 2, 1568–1581. Samoilenko, S. and Osei-Bryson, K.M. (2014). Formulation of Context-Dependent and Target-Specific Strategies of the Impacts of ICT on Development. In Proceedings of the 7th Annual SIG GlobDev Pre-ICIS Workshop ICT in Global Development, Auckland, New Zealand, December 14, 2014.

22

Discovering Common Causal Structures that Describe Context-Diverse Heterogeneous Groups

INTRODUCTION It is commonly acknowledged that operational excellence is one of the sources of competitive advantage of modern enterprises. Fundamentally, the concept refers to achieving a high level of efficiency of conversion of inputs into outputs, where a higher level of efficiency implies, ceteris paribus, a greater degree of excellence. Modern organizations typically regard information and communication technologies (ICTs) as one of the significant direct or indirect inputs for achieving such operational excellence and competitive advantage. Since the concept of competitive advantage involves a relative comparison of the performance of organizational entities, the concepts of organizational capabilities and benchmarking (e.g., Gouveia, Dias, Antunes, Boucinha & Inácio, 2015) are relevant. Ayabakan, Bardhan & Zheng (2017) noted that with regards to the investigation of this pair of concepts: “A dominant approach in IS research involves the use of survey instruments designed to elicit user responses on their perceptions about competencies and capabilities … A limitation of such perception-based approaches is that they represent a subjective measure of firm/organizational capabilities”; these researchers therefore proposed an approach that involves the use of non-subjective data and the data envelopment analysis (DEA) method for doing benchmarking (e.g., Adler, Liebert & Yazhemsky, 2013; LaPlante & Paradi, 2015) of the Input–Output conversion process. In this chapter, we not only take a similar approach to benchmarking but are also interested in the context of the Input–Output conversion process, including those that involve ICTs as input(s). The motivating idea for this research project is that the process of benchmarking could possibly be enhanced by the discovery and application of non-obvious common causal structures that differentiate more efficient organizational entities from less efficient ones. This motivating idea triggered our intention to design an appropriate methodology artifact that involves the analysis of nonsubjective data. This research project can be considered to fall within the realm of information systems (IS) research for at least the following reasons: (1) benchmarking research is an aspect of well-established IS/ICT & Productivity research stream (e.g., Hitt & Brynjolfsson, 1996; Ko & Osei-Bryson, 2004); (2) benchmarking research 157

158

Quantitative Methodologies using Multi-Methods

has appeared in leading IS journals (e.g., Ayabakan, Bardhan & Zheng, 2017); (3) the proposed solution artifact involves the a creative integration of several IS artifacts (i.e., multiple data mining methods) with DEA; and (4) our illustrative example falls within the well-established IS/ICT & Productivity research stream.

A CONCEPTUALIZATION OF THE BENCHMARKING PROBLEM The term “benchmarking” is a commonly used one, and popular terms tend to be vulnerable to falling prey to the unfortunate assumption of the universality of their meaning. The concept of benchmarking is important to our inquiry; thus, we feel it is warranted if we spend a few sentences making sure that we clarify the chosen meaning of the term to our readers. Fundamentally and historically, benchmarking means accurate application of a measure, whatever the measure of interest could be (Merriam-Webster, 2017). Consequently, benchmarking is inherently a two-part process. Firstly, the presence of some sort of a standard measure, or a target, is established. And secondly, the process of emulation of that target is undertaken. For example, when we use a ruler to draw a 10-inch line on a piece of a chapter, we are benchmarking a chosen target (e.g., 10 inches) in the context of the piece of a chapter. We can generalize benchmarking as a process of emulation of the target in a new context. This, immediately, brings up the problem of representation of the target, for in order to emulate a target we need to know the attributes that sufficiently describe (e.g., represent) the target. This is not an easy undertaking in the case of complex targets. One thing is for a freshman to benchmark her academic performance by aiming to have 4.0 GPA (a trivial one-dimensional construct), and another thing for her to benchmark against the likes of Einstein (a non-trivial multi-dimensional construct). Similarly, it is easy for a competitor to benchmark the battery life of cellular phone of the industry leader, and it is hard to benchmark the phone itself. The whole concept of formal representation, which is relevant to benchmarking specifically and underlies the whole field of computing generally (e.g., if it can be formally represented, then it can be computed), is dependent on two factors. One is objectivity and the other is scope. The factor of objectivity of representation refers to having an objective (e.g., standard, agreed upon) scale for a given characteristic of interest, where an attribute “Sugar Content” could be objectively represented via “grams per kilogram” scale, and subjectively so via scale “perceived sweetness”. The factor of objectivity, which is dealt with by finding or creating an appropriate measurement scale, is much easier to address than the factor of scope. In simple terms, scope deals with selecting what is “in” and what is “out”, deciding on a set of attributes that adequately model (e.g., describe, represent) the target. The complexity of the decision is directly related to the complexity of the target; consequently, it is easier to benchmark a body mass index (BMI) than a luxury car, or a successful firm. This issue of scope is not a trivial one because from a complex systems perspective we don’t have a philosophical basis for making the decision regarding what the adequate description of the system itself would be. While a component-based description seems to be good and easy beginning things get increasingly complicated once we start considering non-linearity of the relationships between the components, various dependencies, and emergent properties. Furthermore, in the case

Discovering Common Causal Structures

159

of complex adaptive systems (e.g., person, organization, or economy) we cannot completely “abstract away” the system’s environment, which further complicates the undertaking of nice and neat scoping of the target. Under such circumstances, the scope of the target of benchmarking is not given, but is a result of an active discovery, and it is in this area that this chapter aims to make a contribution. While the concept of benchmarking can be applied to a great variety of contexts, in this chapter we apply it to economic units, which we define as a set of entities (e.g., firm, economy) that: (a) transform a set of expense-associated inputs into a set of revenue-associated outputs and (b) aim to minimize expenses and maximize revenues. Any viable economic entity aims to ensure its survival by means of adopting and maintaining a valid business model. One of the purposes of a business model is to ensure that the stream of revenues is greater than the stream of expenses. While a valid business model assures an operational-level day-to-day viability of the entity, it does not guarantee long-term survival, for an entity could be doomed due to failed strategic- and tactical-level initiatives, such as poorly chosen and implemented strategy, erroneous vision, or misguided business goals. In order for a business model to bring the intended results, it must be implemented – this is done via business processes. For example, a business model “sale of product to customers” in the context of a bakery can be implemented by means of “acquire inventory”, “bake products”, and “sell baked goods” processes. It is an effective and efficient execution of business processes that constitutes the operational excellence of firms, and it is not surprising, therefore, that the less successful firms often aim to improve their operational performance by means of benchmarking of the business processes of their more successful counterparts. Our focus in this research is on benchmarking of business processes within the context of economic units, and, conceptually, the problem that we are trying to help addressing is associated with the necessarily different contexts of the target and the destination of the benchmarking. Let us consider two bakeries – one being a target of benchmarking (e.g., highly efficient) and another one being a destination of benchmarking (e.g., less efficient). Clearly, the process of baking is important and is easy to define based on the “ingredients → baked products” model. However, the same process takes place within different contexts: two bakeries may have a different number of employees with a different number of years of experience, they may have different equipment, and they may have different environmental conditions. Thus, the context of the process of baking is also important but is difficult to define, for there is no common context model for the target and the destination. Conceptually, the question can be expressed as: How to scope the target of benchmarking so it takes into consideration contextual factors that are relevant and applicable to the destination of the benchmarking?

RESEARCH PROBLEM AND RESEARCH QUESTIONS OF THE STUDY Fundamentally, any business process can be seen as a process of conversion of means into ends, where the primary goal is to minimize the cost of means and maximize the value of ends via increasing efficiency and effectiveness of the mechanism

160

Quantitative Methodologies using Multi-Methods

of transformation. If we consider a concept of a business process from a structural perspective, we can identify three distinctive parts. First, there is a set of inputs, second, there is a set of outputs, and third, there is a mechanism of transformation of inputs into outputs. This structural decomposition of a business process is important because it allows identifying a component, or components, that are relevant to the process of benchmarking. A simplified model of conversion of inputs into outputs will take into consideration only those inputs that are necessary for producing the required output. Such a “pure” Input–Output process is akin to a recipe, which specifies a set of ingredients required to make a dish. Once an optimal recipe is developed and acknowledged as the model to be emulated, be it for making pies, or fragrances, or cars, or cellular phones, it is becoming fixed and pretty much ready for benchmarking. There are, of course, well-known exceptions of “secret recipes” of Coca-Cola, Pepsi, KFC, Bush’s, and McDonald, but those are what they are – exceptions to the rule of what otherwise can be called recipe transparency. Let us consider an example of Company A trying to benchmark the process of production of Apple’s iPhone 8. The output of the process is known – it is the phone itself. The inputs are also known – the teardown (e.g., Techinsights.com, 2018) provides a complete list of the inputs – required components. Furthermore, the process of conversion of inputs into outputs is also known – there is a specific way in which parts are assembled into the whole. Would Company A, knowing all the inputs, knowing the mechanism of transformation (assembly), and knowing the output, succeed in getting the same result as Apple have? The answer is simple: it depends on the context – in a perfect world, yes, and in a real world, no. It is the context that presents most of the problems, and the question is how to account for the differences in the contexts. We suggest that it can be done by means of expanding a set of inputs and associated outputs beyond what is required by the pure production process and incorporating relevant context-reflecting factors. In addition to context-specific factors, a more realistic process of conversion of means into ends would also take into consideration inputs and outputs of a conceptual nature represented via proxies (e.g., “level of customer satisfaction” represented via “# of repeat purchases”). This is especially true for enterprises that must consider not only pure inputs and outputs of a production process but also a specificity of the context within which inputs are obtained and utilized, as well as the context for which the outputs are produced – outcomes of the process of conversion of inputs and outputs. Granted, a representation of a pure, chemistry lab-like, input–output process of conversion may serve well in push-dominated industries. However, one of the outcomes of globalization is a more prevalent pull-dominated marketplace that forces consideration of the context – where the ease of access to inputs and desirability of outputs are the intuitive factors that come to mind. An enterprise that makes bricks and an enterprise that specializes in fast fashion may both rely on operational excellence, but the two will differ greatly in regard to what, conceptually, constitutes a set of inputs and a set of outputs of their respective operations. Even enterprises of the same type within the same industry (and even if owned by the same company (Korotin, Popov, Tolokonsky, Islamov & Ulchenkov, 2017) within the same industry) often require consideration of the context within which

Discovering Common Causal Structures

161

their operations take place. For example, a meaningful comparison of the efficiency of an oil rig in Siberia with an oil rig in Saudi Arabia must take into consideration the contexts within which rigs operate. Similarly, a measure of efficiency of a delivery company should take customer satisfaction into consideration by using some sort of a proxy on the output side of the conversion equation, for an “amount of fuel → distance travelled” model is of limited use. Consequently, firms must be looking for a model of the type of “Input→Output→Outcome”, which, in the case of a delivery company would be expressed as “Amount of Fuel→Distance Travelled →Customer Satisfaction”. In both cases, one would have to consider variables that are not endemic to the Input–Output process, but rather reflective of the environment within which the process takes place. Keeping this in mind we advance the following proposition: An assessment of the level of operational excellence of an enterprise operating in pull-driven global marketplace must take into consideration the specificity of the context of the conversion of Inputs into Outputs as well as a measure of outcome of the process. Once we start considering the relevance of the context and outcomes of the process of Input–Output conversion to the idea of operational excellence we arrive at two intuitive implications. First, the possible domain of Input–Output attributes expands beyond the actual, “true” variables involved in the process. While a compounding pharmacy may rely on a “true” process of conversion of the “true” Inputs (i.e., chemical components), into “true” Outputs (i.e., drugs), an outlet mall will have a fundamentally different perspective on what the “true” process is, and what are the Inputs and Outputs of the process are. Second, commonly utilized process management and process improvement techniques (e.g., Kaizen, Six Sigma, etc.) become insufficient due to their intra-process and output-oriented foci. One of the options to improve the level of operational excellence is by benchmarking (adoption of Enterprise Systems falls under this category), where a company seeks a worthy-to-emulate counterpart (within, or outside of one’s industry). Another option is to re-engineer the existing process (e.g., via Business Process Re-engineering). Both approaches have been utilized, and both approaches have been criticized (Al-Mashari et al., 2001; Fui-Hoon Nah et al., 2001; Cao, Clarke, and Lehaney, 2001; Moon, 2008). In the case of Business Process Re-engineering (BPR), there is too much emphasis on the context with the goal of creating unique processes, and in the case of benchmarking via Enterprise Systems (ES), there is too little emphasis on the context of the existing business processes due to their replacement by the inherent processes embedded in the ES. We suggest that the process of benchmarking could be enhanced by discovering non-obvious common causal structures that may be shared by the involved entities. Let us consider a distribution company that operates multiple warehouses in different states in the US. And let us suppose that we are interested in the process that we will call “Receiving a Shipment from a Manufacturer → Delivering the Product to the Department Store”. Fundamentally, the process of increasing a level of efficiency relies on two insights: first, the insight regarding the process itself – the way the shipment is received, then stored, then accessed, and then shipped out. This is a

162

Quantitative Methodologies using Multi-Methods

straightforward optimization problem that could be tackled by century-old methods of Frederick Taylor (1914). The second insight, however, is more important, for it deals with the context of the process – what are the specific conditions that impact the process, and what are the conditions that differentiate less efficient warehouses from more efficient ones. Some of the conditions are uncontrollable – a crew running an oil rig in Siberia cannot do much about an insight “outside temperature of greater than 85 degrees → higher oil production” received from their counterpart in Saudi Arabia. However, some of the insights could be utilized – an insight “average experience of a crew member is greater than 5 years → higher oil production” is an actionable one. This leads us to a second major proposition: The impact of the process of benchmarking that is geared towards increasing the level of operational excellence could be enhanced if it is augmented by the actionable insights, in the form of non-obvious causal structures, regarding the context within which the operations take place. At this point, we would like to illustrate the meaning of the concept of “non-obvious causal structures” to our reader. First, we define the concept of “causal structure” as any type of “if A, then B”. An “obvious causal structure”, accordingly, is a recipelike depiction of a production process, exemplified by an equation for a commonly utilized concrete mix: “4 parts crushed rock + 2 parts sand + 1 part cement → 7 parts concrete mix”. Let us consider Table 22.1. The obvious meta-causal structure for both stores is in the form “Investments → Revenues”, where if Store A operates under the model “High_Investments → High_Revenues”, then Store B operates under the model “Low_Investments → Low_Revenues”. The simple solution for Store B is to change the model to “High_Investments” and hope that “High_Revenues” will transpire on

TABLE 22.1 Contextual Translation of Meta-Causal Structures into Specific Structures Context Meta Input Specific Input Generic Model Investments, in $ Inp_Target 1 Inp_Target 2, etc. Store A: High High Level of Hiring more staff, operational Investments Staff Training, efficiency, Store Maintenance, High-profit InventoryType_1, ratio InventoryType_2, etc. Store B: Low Low Level of Hiring more staff, operational Investments Staff Training, efficiency, Store Maintenance, Low profit InventoryType_1, ratio InventoryType_2, etc.

Specific Output Meta Output Out_Target 1, Revenues, in $ Out_Target 2, etc. Number of Associates, High Level of Customer Satisfaction, Revenues Sales_InventoryType_1, Sales_InventoryType_2, etc. Number of Associates, Low Level of Customer Satisfaction, Revenues Sales_InventoryType_1, Sales_InventoryType_2, etc.

Discovering Common Causal Structures

163

the right side. However, investments must be allocated toward something (see Specific Input column) – the question is how to allocate the money in the most beneficial way. If Store B benchmarks Store A, then certain causal structures may not be applicable. For example, due to the difference in climate “InventoryType_1 → Sales_InventoryType_1” model may work wonderfully well for Store A, but would not be applicable to Store B. An obvious causal structure “Hiring more staff → Number of Associates” will always work, but it is not clear if this is the best bang for a buck for Store B. A set of possible non-obvious causal structures could be depicted by the model “Hiring more staff, Staff Training, Store Maintenance → Customer Satisfaction”. Given limited resources of Store B, then to which one of the target inputs should investments be allocated? The question is simple to answer if we discover that Store A and Store B do not differ greatly in terms of number of associates, or money spent on store maintenance, but do differ significantly in terms of the money spent on staff training. Consequently, under this scenario a non-obvious causal structure would be “Staff Training → Customer Satisfaction”, and if this model is currently implemented as “Staff Training= High → Customer Satisfaction=High” in Store A, and as “Staff Training=Low → Customer Satisfaction=Low”, then Store B could consider investing more money in this area knowing that it worked in Store A. Let us consider yet another example of a company that operates a set of similar retail stores in various locations in the US. At the very high level, the process that is shared by all the stores is “Investments in the store → Revenues from the store”, but this process takes place in different contexts. In such cases, the term “context” means not only geographic location, climate, average income of the local population, etc. but also includes details regarding the application of investments. Meaning, when a certain amount of money is allocated, it is eventually allocated toward something – renovation of the store, upgrade of the equipment, and training of the staff. Furthermore, investments could be allocated toward purchasing different types of inventory – in some cases, it could be sporting goods, in other cases, it could be jewelry, or it could be household items. A less efficient store in Arizona may not benefit much from the insight “Increase in Inventory of Down Jackets → Increase in Revenue” offered by a more efficient store in Alaska, but it could benefit from an insight “hours of annual staff training >20 → increase in revenue”. The trick then becomes is how to uncover the actionable insights that allow for increasing the level of operational excellence. All the points made above could be summarized in the following set of propositions: 1. Less efficient enterprises could improve their performance via benchmarking their more efficient counterparts (a Minor League baseball pitcher could benchmark a Major League pitcher) 2. The difference in the level of efficiency is due to two factors – efficiency of the process of conversion of Inputs into Outputs, and the context within which the process takes place (a Major League pitcher could be better than a Minor league pitcher because of the superior technique – the process of pitching, or due to the context within which the process is utilized – the experience, strength, height, hours spent in the gym, etc.)

164

Quantitative Methodologies using Multi-Methods

3. Some of the contextual factors that impact the process are shared by less and more efficient enterprises are controllable – by improving these factors less efficient enterprises can increase the level of efficiency of the Input– Output Process (while a Minor League pitcher cannot become taller, he can increase his level of efficiency by spending more time in the gym) 4. Discovering a set of controllable contextual factors would enhance the result of a benchmarking, thus allowing less efficient enterprises to increase their level of efficiency. Taking into consideration these examples and propositions, at this point, we can outline the general problem that this chapter deals with, namely: How to discover a common set of non-obvious causal structures that differentiate less and more efficient entities? We can also identify a population for which a solution to the problem could be of benefit, which is a: A sample of convenience comprised of efficiency-driven entities, or groups of entities, that could be described by the same “Input → Output” production model and the same set of contextual attributes. Arguably, the process of benchmarking with the purpose of increasing the level of efficiency could be viewed through the lens of hypothetico-deductive logic (Godfrey-Smith, 2003) for, fundamentally, the process of benchmarking is causally deterministic in nature (e.g., impact C to get D), conclusion driven (e.g., change from Conclusion: Inefficient to Conclusion: Efficient), and based on the attempt to alter the state of minor premise to impact the conclusion. This perspective is illustrated in Table 22.2. TABLE 22.2 Example of Hypothetico-Deductive Logic (HDL) Perspective on Benchmarking Steps of HDL Major Premise

General Example Efficient entity N has a causal structure A with a state X

Minor Premise

Inefficient entity M has a causal structure A with a state Y

Conclusion

Efficiency of Entity M could be impacted by the state Y of A Entity M may benefit from changing the state of structure A from Y to X

Insight

Example of “Training → Reorders” Causal Structure Efficient sales team N annually spends $700/ salesperson on training and has 50 reorders per salesperson Inefficient sales team M annually spends $200/salesperson on training and has 10 reorders per salesperson Efficiency of team M could be impacted by low spending on training Team M may benefit by spending more on training

Discovering Common Causal Structures

165

The importance of solving the general problem of this study is intuitive, for: Identifying a set of common causal structures that differentiate less and more efficient entities allows for obtaining insights regarding the ways of impacting the level of efficiency. As our reader can see, the purpose of this investigation is to find a way of discovering, basically, major and minor premises of the steps of hypothetico-deductive method. Let us offer the following illustrative example: Major premise: A functionality (Input–Output process) N with the state A indicates relative efficiency. Minor premise: Entity X has a functionality N with the state B. Conclusion: Entity X is relatively inefficient. Problem: Entity X must become relatively efficient – change Conclusion of HDL. Benchmarking solution to the Problem: Entity X may change Conclusion by changing Minor premise – via matching Minor premise to Major premise and obtaining functionality N with the state A, not B. Consequently, we put forward a set of propositions, expressed as follows: Proposition 1.1: The Process of benchmarking relies on the rules of hypothetico-deductive logic. Justification: The process of benchmarking relies on matching the actual state (Minor premise) of an entity with the desired state (Major premise) of an entity with the intent of changing the resultant behavior of the entity (Conclusion). Proposition 1.2: The problem of process inefficiency could be addressed by benchmarking. Justification: The state of an entity could be depicted by the process of conversion of its Inputs into Outputs. Proposition 1.3: Success in addressing the problem of process inefficiency via benchmarking is impacted by the discovery of a proper model of conversion of Inputs into Outputs. Justification: A model of conversion of Inputs into Outputs is not limited to pure Inputs and Outputs, but must also take into consideration relevant contextual factors. Proposition 1.4: A proper model of conversion of Inputs into Outputs could be expressed as a combination of two models – the “true” process model (dictated by the “true” production process, or given by the subject-matter expert), and the “contextual” process model (to be discovered in the case of context-dependency of the “true” process) that impacts, directly or indirectly, the “true” model. Justification: An enterprise is an open dynamic complex system that impacts, and is impacted by its context – competitive business environment (Amagoh, 2008; Samoilenko & Osei-Bryson, 2007a; Samoilenko, 2008a).

166

Quantitative Methodologies using Multi-Methods

Proposition 1.5: The overall success of the benchmarking effort is impacted by the discovery of a relevant contextual process model which could be expressed in the form of non-obvious causal structures. Justification: If a context impacts the process of conversion of Inputs into Outputs, then it is important to depict the context properly. To formalize the process of discovery of causal structures we express our inquiry in a form of a methodology – this allows our readers to formally review and evaluate our approach. The dynamic nature of the global business environment implies that as the time goes by and the competitive landscape changes the membership of the sample will also change and the causal structures will have to be discovered anew. Thus, we offer our solution to the stated above problem in the form of inductive methodology relying on non-parametric data analytic techniques and methods. Conceptually, we present the scope of the research problem that our methodology addresses in Figure 22.1. Prior to describing our methodology, we would like to present to our reader a conceptual argument for why the proposed methodology should exist at all. A simple way of doing it is by presenting a set of propositions so that our readers could follow along the logic of our argument in a step-by-step fashion. Proposition 2.1: Deductive approaches to process improvement (e.g., benchmarking) relies on the bounded and well-defined universe of discourse (UoD). Proposition 2.2: Deductive approaches to process improvement are based on the rules of Hypothetico-Deductive Logic (HDL). Proposition 2.3: Bounded and well-defined UoD is expressed as a set of Major and Minor premises of HDL, alteration of which results in changes to Conclusion of HDL. Proposition 2.4: Changes in the dynamic business environment force redefinition of the UoD. Proposition 2.5: In order to apply hypothetico-deductive approaches to process improvement in the context of changing business environment a new model of UoD must be created, re-bounded, and re-defined.

FIGURE 22.1  Scope of the research problem.

Discovering Common Causal Structures

167

Proposition 2.6: A new model of UoD reflecting a new business environment cannot be created using deductive (general → specific, current → past) methods, but can be created using inductive (specific → general, past → future) methods. Proposition 2.7: Data mining and non-parametric methods allow for creating the new model of UoD via identifying changes that took place in the business environment. Overall, this set of propositions could be visually expressed in the form of a Methodological Spiral depicted in Figure 22.2 which outlines the scope of our inquiry, and to set the limits of what our methodology can do. We summarize the applicability of our approach as follows: Limitation 1: Our methodology is not applicable to static business environments, for such environment will not require re-defining UoD. Our methodology will be of greater use to firms that develop on-line services and provide outsourced customer support, that to the firms that build airports, oil pipelines, and chip manufacturing facilities. Limitation 2: Our methodology is not applicable to the pure production process that is context –independent – not impacted by the business environment. For example, a process of manufacturing of Advil according to the given formula will not be aided by our methodology, while a process of a delivery of on-line orders will be. Limitation 3: Our methodology is of limited importance to the process improvement approaches that are inductive in nature. Fundamentally, deductive approaches are based on copying of the best practices in changing contexts, and our methodology can offer valuable insights regarding changes that took place. Inductive approaches, such as business process re-engineering, are not based on copying of the best practices, but rather on creating unique best practices.

FIGURE 22.2  Methodological spiral as a response to the changes in business environment.

168

Quantitative Methodologies using Multi-Methods

FIGURE 22.3  Benchmarking as a generalized solution to the problem of process inefficiency.

Keeping this in mind, we can generalize the solution to the problem of process inefficiency via benchmarking as consisting of two elements. First, addressing process-specific factors, and, second, addressing context-related factors. The proposed solution is presented in Figure 22.3. The problem of relative inefficiency could be tackled using a two-step approach. The first step deals with identifying relatively less-efficient and relatively moreefficient counterparts (e.g., Leaders vs. Followers). The second step deals with inquiring into the factors that could be associated with the relative inefficiency and then implementing changes to the factors. This approach works well if the scope of the undertaking is limited to the input–output conversion process. Sometimes, however, the roles are assigned based on external to the input–output process’ factors, when, for example, less affluent economies try to benchmark their wealthier counterparts. In any case, the evaluation of the efficiency of the process of the conversion of inputs into outputs takes place.

THE PROPOSED METHODOLOGY We now present a description of the phases of the methodology followed by its justification.

Description of the Methodology Phase 1 Define the Transformation Framework

Description Such a framework would involve the specification of the relevant Driver and Impact constructs and their indicator variables. It would form the basis for identifying the potential causal paths that are to be evaluated. An existing established transformation framework (e.g., Networking Readiness Index) could be utilized, or a contextrelevant model could be developed using logical informed arguments & insights from relevant existing Literature

Discovering Common Causal Structures Phase 2 Partition the set of Decision Making Units into Meaningful Groups

3

Benchmark the DMUs

4

Identify the Major Attributes that Differentiate the DMUs

5

Discover the Causal Structures

169

Description This may be done using a previously established rules for group membership (e.g., World Bank classification scheme), or the application of a data mining-based clustering process on a data set includes the Input (i.e., Drivers), Mediator and Output (i.e., Impacts) variables of the Transformation Framework For each group: For each potential Causal Path, apply the DEA method to the data set in order to identify classify the DMUs in each DMU group in terms of the following: (a) Efficient vs. Non-Efficient; and (b) Growth in Productivity vs. Non-Growth in Productivity. For each group: • Apply Decision Tree Induction (DTI) in order to determine the major (i.e., top) attributes that determine the status of each DMU in terms of Relative Efficiency. • Apply DTI in order to determine the top attributes that determine the status of each DMU in terms of Growth in Productivity. For each group: • Apply Association Rules Mining (ARM) to discover the significant rules that describe the causal structures that are associated with: (a) Efficient vs. Non-Efficient; and (b) Growth in Productivity vs. Non-Growth in Productivity.

Justification & Benefits of the Methodology The first phase of our methodology is associated with defining, or adapting, a context-specific transformation framework according to which inputs are converted into outputs. For all intents and purposes, the goal of the transformation framework is to create a set of established pathways by which the transformation of inputs into outputs takes place. We refer to the transformation framework as the Universe of Discourse of Transformation (UoDoT) – an established common set of means by which inputs could be transformed into outputs. A meaningful for our purposes UoDoT is semi-constrained in the sense that it exists between completely constrained and unconstrained states. Let us consider an illustrative example. Let us consider Joe and Bob, two people who have some money to invest with the purpose of receiving a profit. So they do so and Joe does better than Bob. The basic high-level model for Joe and Bob is “Money → Investment → Profit” and the basic goal is the maximization of the outputs – profit. The first and the last parts of equation, money and profit, are perfectly clear – Joe and Bob have money, and they want to get more money. The problematic part is the second one – that of investment, because until Joe and Bob know what, exactly, it stands for, they cannot optimize the input–output process of “money → profit”. And it is investment part in this example that we refer to as UoDoT, and it deals with the question of what is it, exactly, Joe and Bob could invest their money to get, as a result, more money. In this example, Bob does not do as well as Joe, and Bob asks for an advice – turns out all Joe does, just as Bob does, is to keep his money on his savings account, and Joe does better because his bank has better

170

Quantitative Methodologies using Multi-Methods

rates than Bob’s bank. This is an example of constrained UoDoT – a single path (e.g., savings account) according to which money transformed into more money – revenue. Conversely, let us consider Joe not sharing his investment strategies with Bob–Joe could be investing in illegal drug trade, or in an underground amphetamine lab, or he could be really lucky playing state lottery, or Joe could be doing million other things. This is an example of unconstrained UoDoT – many different paths could exist, and no set of paths is known as the actual one. As our reader can see constrained UoDoT and unconstrained UoDoT do not allow for benchmarking to take place – in the former case there is nothing to benchmark, and in the latter case it is not known what to benchmark. The situation changes, however, if we have a semi-constrained UoDoT where Joe invests his money in bonds, stocks, and hedge funds. In this case, the basic high-level model for Joe and Bob is “money → {bonds, stocks, and hedge funds} → profit”, and Bob can benchmark Joe by learning intricacies of his investment strategies – when and how much to invest into what and under what conditions. It seems like the Joe and Bob scenario works fine for the illustrative purposes, but it leaves out one important point, namely, how does Bob know that Joe does better than him? Or, putting it in a general way: how does an entity identifies its benchmarking target? The second part of our methodology is designed to provide an answer to this question. It is during the second phase a data sample is partitioned into various sub-groups with the purpose of identifying a better performing entity or sub-group (benchmarking target), and a worse performing entity or sub-group, and various levels of entities or sub-groups in between. This could be done according to a given scheme (e.g., classification of IMF based on the income level of economies), or expert knowledge (e.g., general manager identifies top-, medium-, and low-performance retail stores), or by discovery via data mining methods. Needless to say, all the members of the sample must share the same semi-constrained UoDoT. DEA is a method that is widely used for the purposes of calculating scores of the relative efficiency of entities that receive inputs and produce outputs. For example, we could compare three groups of basketball players of different levels (e.g., high school, college, and professional) in terms of their efficiency of conversion of minutes played and attempts taken into assists and points. Results of DEA would yield the most relatively efficient group, but because DEA model is a “black box” model we would not know what differentiates the groups, or why one group is more efficient than the other two. The insights could be provided by Decision Tree Induction (DTI), which would yield an attribute, or a few attributes, that differentiate the groups. Thus, given a set of attributes describing the three groups of basketball players, we may find out that the main difference between the groups is in years of experience and hours of weekly practice. However, there are plenty of players with many years of experience who train long hours every week, but don’t play so well. We would like to know what set of attributes is actually associated with the outputs – assists or points scored. Here is where ARM may help. ARM allows for generating sets of association rules of the type “If (a,b,c) Then (d)”. This is very valuable, for we can see the patterns of associations specific to each group. However, ARM tends to generate many rules, some trivial, some meaningless/ non-actionable, and some useful. The problem with selecting the rules describing the different groups is that the rules may contain completely different attributes – this

Discovering Common Causal Structures

171

would result in comparing apples and oranges. For example, in the case of basketball players we may get “If (height > x) Then (minutes_played > y” for high school players, “If (experience > n) Then (assists > m)” and so on. So, the trick is to identify a set of rules that is based on a set of common criteria that differentiates the groups – this insight is provided by DTI. Consequently, the novelty of our approach is associated with its capability to identify the main differentiating factors responsible for heterogeneity of the context, and then to base the selection of the rules on those factors. Previous investigations used a hybrid DEA/DTI methodology (Samoilenko & Osei-Bryson, 2008), and the use of ARM with DEA was recently reported by Samoilenko (2016); however, this investigation represents the first case of using the three methods (i.e., DEA, DTI, ARM) in synergy. Simply put, if DEA allows us to identify the efficient performers, and DTI helps us to discover the relevant dimensions that differentiate efficient and inefficient performers, then ARM allows us to benchmark efficient performers via a set of “If → Then” rules that rely on the dimensions discovered by DTI analysis. To our knowledge, no other combination of data analytic and data mining methods could offer so much in so few steps. Banker & Natarajan (2008) had also presented a 2-stage DEA+OLS method which assumes that contextual variables have been identified but they provided no guidance as to how such variables could be identified. Appropriate application of Banker & Natarajan’s and similar methods in a real-world context would thus require that before such methods are applied that appropriate activities to factor in context be conducted. Phases 1 & 2 of our proposed framework describe such a set of activities. Worthington & Dollery (2002) had suggested that “environmental (or contextual) factors may encompass both physical environmental circumstances, as well as constraints arising from organizational and managerial policies”. Later Barbosa, Lima & Brusca (2016) noted that: “the scores of efficiency are calculated from a set of inputs and outputs; however, the explanatory variables of these models do not represent inputs or outputs, but the context where each DMU is”. In this chapter, we consider the concept of context in a manner similar to that of Worthington & Dollery (2002) and Barbosa, Lima & Brusca (2016); we consider the context to involve non-discretionary explanatory variables that are not used as inputs in the calculation of the efficiency scores; contextual variables could be categorical, ordinal or interval. Thus, for example, given a heterogeneous set of DMUs, the group to which a given DMU belongs is an aspect of the context in which it operates as it converts its inputs to outputs. Since as noted by Worthington & Dollery (2002) and other researchers that ignoring context “may lead to disingenuous efficiency measures” we present a multi-phase framework that that aims to provide guidance for the identification of the aspects of context before applying DEA. Banker & Natarajan (2008) demonstrated that under a specific set of conditions (e.g., “monotone increasing and concave production function separable from a parametric function of the contextual variables”) a 2-stage DEA+OLS model yields “consistent estimators of the impact of the contextual variables”. It is not clear if such conditions typically apply in real-world contexts, but when they do apply then their 2-stage DEA+OLS model could be applied in what corresponds to Phases 3 & 4 of our framework to identify the variables that impact productivity. Interestingly, the use of DTI to replace OLS in the stage model does not require the pre-requisite conditions assumed by Banker & Natarajan (2008).

172

Quantitative Methodologies using Multi-Methods

ILLUSTRATIVE EXAMPLE – APPLICATION TO SUB-SAHARAN ECONOMIES Our illustrative example involves the application of the methodology to a data set that covers a sub-set of the countries of Sub-Saharan Africa (SSA). We obtained the data from a publicly available source – World Economic Forum’ Global Information Technology Report (GITR, 2015). In 2012 the representation of the NRI has partially changed, so we decided to concentrate on the new version of NRI and use the data for 2012–2015 period. However, for some SSA economies data was missing for some years (e.g., Angola, Seychelles, Liberia, Gabon, Sierra Leone, and Guinea) and represents 27 SSA economies.

Phase 1: Define the Transformation Framework In our illustrative example, we will use the NRI framework (Dutta, Geiger & Lanvin, 2015), the adapted version of which is depicted below in Figure 22.4. The framework relies on four subindexes and their ten sub-categories (or pillars) to obtain the value of the NRI, which reflects the capacity of economies to benefit from ICT. An increase in the value of NRI for a given economy is indicative of the increase in the impact of ICT on innovation and productivity (Dutta & Jain, 2003). Interestingly, the original framework does not explicitly connect Environment, Readiness, and Usage subindexes (referred to as Drivers within the framework) with Impact subindex (referred to as Impact), despite relying on a principle that “…the environment, readiness, and use—interact, co-evolve, and reinforce each other to create greater impact” (Di Battista, Dutta, Geiger & Lanvin, 2015, p.4). We scope our inquiry by only considering relationships between Drivers (environment, readiness, and use) and Impact (socio-economic impact of Drivers) – as indicated by arrows in Figure 22.4. All possible interactions within Drivers (e.g., between environment, readiness, and use) we consider to be beyond the scope of our investigation. In our inquiry, we use the framework depicted in Figure 22.4 to

FIGURE 22.4  The framework of networked readiness (adapted format).

Discovering Common Causal Structures

173

investigate, via DEA, the efficiency of the process by which Environment, Readiness, and Usage subindexes impact two sub-categories of Impact subindex – Economic and Social impacts.

Phase 2: Partition the Set of Decision Making Units into Meaningful Groups Our set of DMUs are countries of SSA. In some of our previous work (e.g., SelfReference) that involved the transition economies (TEs) of Eastern Europe, we used cluster analysis to partition the set of TEs into two meaningful groups. While that option is available, so are other options include World Bank classification schemes. So, in this study, we will use the classification of the World Bank as of July 2014 to partition the set of SSAs into three groups: Low Income, Low-Middle Income & Upper-Middle Income (see Table 22.3). We obtained the data from a reputable source – the World Economic Forum’s Global Information Technology Report 2015 (GITR, 2015). In some cases, the representation of SSA economies was inconsistent, for example, we could not include Angola, Seychelles, Liberia, Gabon, Sierra Leone, and Guinea in our sample because the data for some of the years was missing. While there is an advantage to increasing the sample size of a study, there is a price to pay via dealing with missing variables, imputation of values, and additional data preprocessing. After considering the pluses and minuses of “sample size vs. data actually available”, we have assembled a smaller data set that contained no missing data and no outliers, but was as reliable as one could get from a given source. Overall, we were able to compile the data set representing 27 SSA economies (the classification of the International Monetary Fund as of October 2014).

Phase 3: Data Envelopment Analysis During the first phase, we rely on DEA to evaluate relative efficiency of three “Drivers → Impact” paths. We will use variable return to scale (VRS) DEA model to conduct the analysis, for it is reasonable to argue that SSA economies have not yet reached the point of developing a level of ICT infrastructure allowing accruing the benefits yielded by capitalizing on economies of scale.

TABLE 22.3 Sample of Sub-Saharan African (SSA) Economies, by Income Level Income Level Low Income Low-Middle Income Upper-Middle Income

Sub-Saharan Africa (SSA) Economies Burkina Faso, Burundi, Chad, Ethiopia, Gambia, Kenya, Madagascar, Malawi, Mali, Mozambique, Rwanda, Tanzania, Uganda, Zimbabwe Cameroon, Cape Verde, Côte d’Ivoire, Ghana, Lesotho, Nigeria, Senegal, Swaziland, Zambia Botswana, Mauritius, Namibia, South Africa

174

Quantitative Methodologies using Multi-Methods

TABLE 22.4 DEA Models of the Illustrative Example DEA Model Drivers → Overall_Impact (DOI) Drivers → Economic_Impact (DEI) Drivers → Social_Impact (DSI)

Inputs of DEA Model Environment Subindex Readiness Subindex Usage Subindex Environment Subindex Readiness Subindex Usage Subindex Environment Subindex Readiness Subindex Usage Subindex

Outputs of DEA Model Impact Subindex

Economic Impact Sub-category

Social Impact Sub-category

Given a four-years time period, we will run DEA 12 times. Consequently, for each economy in the sample, we are going to have four scores of relative efficiency for each of the three models (see Table 22.4). At this point, we need to provide a justification for the inputs and outputs included in our models (Cook, Tone & Zhu 2014). In regard to outputs, the reasoning is intuitive – first, we would like to assess the efficiency of the overall impact, and then, each type of the impact separately. This is because an economy could be efficient in obtaining one type of an impact (e.g., economic) and not efficient in regard to another impact (e.g., social). With regard to the choice of the inputs of DEA model, our approach is methodological. While we are free to use eight sub-categories of Drivers as inputs of a DEA model, the general rule of thumb is that for a reasonable level of discrimination number of economies (or Decision Making Units in DEA terms) must be at least twice the product of inputs and outputs (Dyson, Allen, Camanho, Podinovski, Sarrico & Shale, 2001). In our case we have a sufficient number of economies in our set, but if we use a DEA model with eight inputs and two outputs then we would need to have at least 2 × 8 × 2 = 32 economies in the sample. Furthermore, and more importantly, the greater the number of factors included in the DEA model, the lower the level of discrimination of the model (Dyson et al., 2001). Our results, summarized in Table 22.5, demonstrated that seven economies of the full sample are relatively efficient with regard to the impact of Drivers on social, economic, and overall impact of ICT. Additionally, we identified relatively efficient economies per each of the incomelevel group; in some cases (e.g., Burundi, Chad, Kenya, Mali, Rwanda, and Senegal), the relatively efficient within its group’ economies are also efficient overall. In other cases (e.g., Swaziland, Lesotho, Botswana, Mauritius, Namibia, and South African Republic), the relatively inefficient, overall, economies end up being efficient within their respective group. Also, we used DEA to calculate the values of Malmquist index (MI), which allows us to assess the changes in the scores of relative efficiency that took over period of time. Under the assumption of constant returns to scale the change indicates changes

175

Discovering Common Causal Structures

TABLE 22.5 Results of DEA Income Level LI

LM

UM

Economy BFA BDI TCD ETH GMB KEN MDG MWI MLI MOZ RWA TZA UGA ZWE CMR CPV GHA NGA SEN SWZ ZMB CIV LSO BWA MUS NAM ZAF

Overall Efficiency of the Impact of ICT Efficient Efficient Efficient Efficient Inefficient Efficient Inefficient Inefficient Efficient Inefficient Efficient Inefficient Inefficient Inefficient Inefficient Inefficient Inefficient Inefficient Efficient Inefficient Inefficient Inefficient Inefficient Inefficient Inefficient Inefficient Inefficient

Overall Changes in Productivity Growth No growth No growth No growth No growth No growth Growth No growth Growth No growth No growth Growth No growth Growth Growth No growth No growth No growth No growth Growth Growth Growth Growth No growth Growth No growth No growth

Growth via EC? Yes Yes No No Yes No Yes No No Yes Yes Yes Yes Yes Yes No No No No Yes Yes Yes Yes No Yes Yes No

via TC? Yes No No No No No No No Yes No No No No No No No No No Yes No No No No No No No No

in productivity. Consequently, we are able to assess whether the economies become more productive or not. Because MI is comprised of two components – change in efficiency (EC) and change in technology (TC), we are also able to assess whether the changes in productivity are associated with a particular component. Overall, only 11 economies (40% of the sample) exhibited growth in productivity. An analysis of the changes in EC and TC offers an interesting insight: 16 economies (60% of the sample) exhibited positive changes in efficiency, but only 3 economies (11% of the sample) demonstrated positive changes in technology. Finally, it is worth noting that only one economy, Burkina Faso, exhibited a balanced growth in productivity, when the growth was driven by both components of MI. Overall, the picture suggests that SSA, as a group, would benefit from a better technology – this suggests that investments in ITC infrastructure should be prioritized.

176

Quantitative Methodologies using Multi-Methods

Finally, it is worth noting that 8 economies (30% of the sample) have not only exhibited a decline in productivity but have exhibited a decline in terms of both components of MI.

Phase 4: Decision Tree Induction (DTI) To proceed with this phase, we need to create a new variable “Target” to differentiate various groups of economies. We are interested in three types of groupings: first, we would like to differentiate the SSA groups by their level of economic development, and then we would like to differentiate the groups in terms of their efficiency of the socioeconomic impact of ICT. The last analysis would involve differentiating SSA economies by growth in productivity – Growth vs. No Growth. Thus, we would conduct DTI three times, which would require Target to have three domains of values. In the first case, grouping by income, the domain of values of Target would be {1, 2, 3}, for, respectively, Low Income (LI), Low Middle Income (LM), and Upper Middle Income (UM) groups of economies. In the second case, grouping by efficiency, The Target would assume the values of {0, 1}, for, respectively, relatively inefficient, and relatively efficient SSA. The same domain of values, namely, {0, 1}, could be applied to the grouping by growth in productivity, where “0” would indicate “No Growth” and “1” would indicate “Growth”. The results of DT analysis summarized in Table 22.6 show that such pillars of NRI as Individual Usage, Business Usage, and Skills Readiness do play important role in differentiating three groups of economies. It is not surprising that there appears to be a clear-cut difference between Low Income and Upper-Middle Income economies, and much less of a difference between Low-Middle Income economies and the other two. We could also identify Individual Usage and Economic Impact as pillars that play role in differentiating relatively efficient SSA economies from inefficient ones. It appears that Infrastructure Readiness, Affordability Readiness, and Individual Usage are factors playing role in differentiating those economies that became more productive from those that didn’t. TABLE 22.6 Results of Decision Tree Analysis Grouping by Economic development

Group Low Income vs. Low Middle Income vs. Upper Middle Income

Relative efficiency

Relatively Efficient vs. Relatively Inefficient Growth vs. No Growth

Change in productivity

Differentiating/Split Variable Individual usage Business usage Skills readiness Individual usage Economic impact Infrastructure readiness Affordability readiness Individual usage Social impact

177

Discovering Common Causal Structures

Phase 5: Association Rule Mining The purpose of this phase is to find possible patterns, associations, or causal structures that may exist in our data. One of the main advantages of association rule mining (ARM) is that it is suitable for undirected data mining; thus, we’ll aim to discover naturally occurring associations between the factors (sub-indexes of Drivers and Impact) – components of NRI. ARM could be classified as either being explanatory or exploratory in nature. In the case of our investigation, we employ exploratory ARM, for we do not have any theoretical support for why certain relationships between the sub-indexes of NRI should exist. A very common approach to generating associations between the variables, or itemsets, via ARM is by using the a priori algorithm (Agrawal & Ramakrishnan, 1994) – we will rely on this approach in the current investigation. Transformation of the data is required for this step – we follow the method of Samoilenko & Osei-Bryson (2017) to do so. The application of ARM result in the generation of multiple association rules (ARs) that contain context-specific combinations of factors (see Table 22.7). The results of the data analysis offered in the previous sections offer evidence that we were successful in developing and testing a novel methodology that provides for

TABLE 22.7 Impact-Specific Rules for Low Income and Low-Middle Income Economies Condition Low Income

Low-Middle Upper-Middle

Low Income, Inefficient

Low-Middle, Efficient

Generated Rules low IND_USE, low BUS_USE low SKILL_READ, low BUS_USE midhigh BUS_USE midhigh SKILL_READ high SKIL_READ, high AFFORD_READ, high GOV_USE high BUS&INNOV_ENV, high INFR_READ, high BUS_USE high BUS&INNOV_ENV, high IND_USE, high BUS_USE low INFR_READ, low BUS_USE low IND_USE, low BUS_USE low BUS_USE, low GOV_USE low BUS_USE, low ECON_IMP

=>

low ECON_IMP

Sup. 21%

Conf. 0.70

Lift 1.9

=>

low ECON_IMP

21%

0.60

1.5

=> => =>

midlow SOCIO_IMP midhigh SOCIO_IMP high SOCIO_IMP

20% 25% 32%

0.50 0.50 1.00

2.5 1.5 2.3

=>

high ECON_IMP

31%

0.85

2.3

=>

high ECON_IMP

31%

0.85

2.2

=>

low SKILL_READ

36%

1.00

1.9

=>

low SKILL_READ

32%

1.00

1.9

=>

low ECON_IMP

25%

1.00

4.0

=>

low SOCIO_IMP

25%

1.00

3.0

178

Quantitative Methodologies using Multi-Methods

investigating complex context-specific relationships between the factors reflecting the state and the impact of ICT Capabilities. The discussion of the results is presented along the points that we considered noteworthy. First, Despite the presence of complex relationships between the Drivers and Impacts of ICT there are common themes associated with the levels of the scores of factors comprising NRI – Business Usage, Individual Usage, and Skills Readiness appear to have a direct relationship with the levels of the scores of socio-economic Impact of ICT. We point out that while the variety of association rules has been generated for a different set of criteria, a common line could also be glanced – some sub-categories of NRI’ subindexes (e.g., related to Skills, Business, Individual usage) appear more frequently than other sub-categories. Second, These results suggest that Business Usage and Individual Usage are among the factors that appear to differentiate economies in terms of their level of economic development, as well as in terms of their relative efficiency of the impact of ICT on the socioeconomic bottom line. These results suggest that wealthier and more efficient economies tend to have higher scores of Business Usage and Individual Usage. The presence of a simple association between the level of income of an economy, its efficiency, and ICT usage seems to be apparent. Third, The Infrastructure and Affordability of ICT seems to have an impact on growth in productivity of SSA economies. This finding is important because it was corroborated, independently, by two different methods of analysis – DEA and DT.

CONCLUSION In this chapter, we have presented a new multi-method methodology for benchmarking that uncovers relevant context-specific causal structures. We then provided an illustrative application of this methodology to an IS/ICT & Productivity research problem in the “developing” countries context that resulted in the uncovering of causal structures that describe relationships between Drivers and Impact of ICT. These results show that the relationships between the factors representing Drivers and Impact are indeed complex – in the sense of Complex Systems Theory where structure of the systems changes (Samoilenko, 2008a) as a response to the pressures of the environment (e.g., only some of the pillars of NRI differentiate the clusters) and the relationships between the components are non-linear (e.g., the association

Discovering Common Causal Structures

179

rules are not stable, but level-dependent). However, the underlying complexity of these relationships can be made more transparent to researchers and practitioners by the application of the proposed methodology. This increase in transparency comes, unsurprisingly, from elucidation of the role of the context within which a complex dynamic system, be it an economy, or an industry, or a firm, operates. While the other commonly used DEA-based methodologies may allow for uncovering the sources of heterogeneity between the various groups in the sample (e.g., Samoilenko & Osei-Bryson, 2007b), or for differentiating on purely efficiency-based reasons for discrepancies in the levels of performance (e.g., Samoilenko & OseiBryson, 2013), our methodology also allows for discovering (unlike the approach of Banker & Natarajan (2008) that takes the “received knowledge” approach) actionable causal structures that are context-based and setting-specific, and in this regard the proposed in this chapter approach is truly unique. In summary, there are several reasons on which claims for the novelty and value of this methodology can be based including: • It subsumes the 2-stage DEA framework of Banker & Natarajan (2008). • It does not require the pre-requisite conditions (e.g., “monotone increasing and concave production function separable from a parametric function of the contextual variables”) assumed by Banker & Natarajan (2008). • It has the capability to identify the main differentiating factors responsible for heterogeneity of the context. • It allows for obtaining actionable information, in the form of non-obvious common causal structures, for improving the performance of the less efficient entities vis-à-vis their more efficient counterparts • It allows for allows for expanding the universe of discourse within which the process improvement initiatives are usually considered, thus allowing to consider the impact of external to the process factors on internal to the process mechanisms. • It involves a new & creative integration of multiple data mining methods (CA, DTI, ARM) with DEA.

ACKNOWLEDGMENT Material in this chapter previously appeared in: A data analytic benchmarking methodology for discovering common causal structures that describe context-diverse heterogeneous groups. Expert Systems with Applications, 117, 330–344.

REFERENCES Adler, N., Liebert, V., and Yazhemsky, E. (2013). Benchmarking Airports from a Managerial Perspective. Omega, 41(2), 442–458, https://doi.org/10.1016/j.omega.2012.02.004. Agrawal, R. and Ramakrishnan S. (1994). Fast Algorithms for Mining Association Rules in Large Databases. In Proceedings of the 20th International Conference on Very Large Data Bases (VLDB ‘94), Jorge B. Bocca, Matthias Jarke, and Carlo Zaniolo (Eds.). Morgan Kaufmann Publishers Inc., San Francisco, CA, 487–499.

180

Quantitative Methodologies using Multi-Methods

Al-Mashari, M., Irani, Z., and Zairi, M. (2001). Business Process Reengineering: A Survey of International Experience. Business Process Management Journal, 7(5), 437–55. Amagoh, F. (2008). Perspectives on Organizational Change: Systems and Complexity Theories. The Innovation Journal: The Public Sector Innovation Journal, 13(3), 1–14. Ayabakan, S., Bardhan, I. R., and Zheng, Z. (2017). A Data Envelopment Analysis Approach to Estimate IT-Enabled Production Capability. MIS Quarterly, 41(1). Banker, R.D. and Natarajan R. (2008). Evaluating Contextual Variables Affecting Productivity Using Data Envelopment Analysis. Operations Research, 56(1), 48–58. Barbosa, A., Lima, S. C. D., and Brusca, I. (2016). Governance and Efficiency in the Brazilian Water Utilities: A Dynamic Analysis in the Process of Universal Access. Utilities Policy, 43(PA), 82–96. Cao, G., Clarke, S., and Lehaney, B. (2001). A Critique of BPR from a Holistic Perspective. Business Process Management Journal, 7 (4), 332–339. Cook, W., Tone, K., and Zhu, J. (2014). Data Envelopment Analysis: Prior to Choosing a Model. Omega, 44, 1–4, https://doi.org/10.1016/j.omega.2013.09.004. Di Battista, A., Dutta, S., Geiger, T., and Lanvin, B. (2015). The Networked Readiness Index 2015: Taking the Pulse of the ICT Revolution. Global Information Technology Report 2015, 3–28. Dutta, S. and Jain, A. (2003). The Networked Readiness of Nations. In The Global Information Technology Report 2002–2003, Dutta, S., A. Lanvin, B., and Paua, F. (eds.). Oxford University Press, New York, NY, Oxford. Dutta, S., Geiger, T., and Lanvin, B. (Eds.) (2015). Global Information Technology Report 2015. ICTs for Inclusive Growth. World Economic Forum and INSEAD, Geneva. Retrieved April 20, 2015 from http://www3.weforum.org/docs/WEF_Global_IT_ Report_2015.pdf Dyson, R.G., Allen R., Camanho A.S., Podinovski, V.V., Sarrico C.S., and Shale, E.A. (2001). Pitfalls and Protocols in DEA. European Journal of Operational Research, 132, pp. 245–259. Fui-Hoon Nah, F., Lee-Shang Lau, J., and Kuang, J. (2001). Critical Factors for Successful Implementation of Enterprise Systems. Business Process Management Journal, 7 (3), 285–296. GITR. (2015). World Economic Forum’ Global Information Technology Report. Available on-line at: http://reports.weforum.org/global-information-technology-report-2015/ network-readiness-index/. Godfrey-Smith, P. (2003). Theory and Reality: An Introduction to the Philosophy of Science. University of Chicago Press. Gouveia, M.C., Dias, L.C., Antunes, C.H., Boucinha, J., and Inácio, C.F. (2015). Benchmarking of Maintenance and Outage Repair in an Electricity Distribution Company using the Value-based DEA Method. Omega, 53, 104–114, https://doi.org/10.1016/j. omega.2014.12.003. Hitt, L. M. and Brynjolfsson, E. (1996). Productivity, Business Profitability, and Consumer Surplus: Three Different Measures of Information Technology Value. MIS Quarterly, 121–142. Ko, M. and Osei-Bryson, K. M. (2004). Using Regression Splines to Assess the Impact of Information Technology Investments on Productivity in the Health Care Industry. Information Systems Journal, 14(1), 43–63. Korotin, V., Popov, V., Tolokonsky, A., Islamov, R., and Ulchenkov, A. (2017). A Multicriteria Approach to Selecting an Optimal Portfolio of Refinery Upgrade Projects under Margin and Tax Regime Uncertainty. Omega, 72, 50–58, https://doi.org/10.1016/j. omega.2016.11.003. LaPlante, A.E. and Paradi, J.C. (2015). Evaluation of Bank Branch Growth Potential using Data Envelopment Analysis. Omega, 52, 33–41, https://doi.org/10.1016/j.omega.2014.10.009.

Discovering Common Causal Structures

181

Merriam-Webster. (2017). Benchmark. Retrieved October 26, 2017, from https://www.merriam-webster.com/dictionary/benchmark Moon, Y.B. (2008). Enterprise Resource Planning (ERP): A Review of the Literature. International Journal of Management and Enterprise Development, 4 (3), 235–64. Samoilenko, S. and Osei-Bryson, K.M. (2007a). Chaos Theory as a Meta-Theoretical Perspective for IS Strategy: Discussion of the Insights and Implications. In Proceedings of the Southern Association for Information Systems Conference, Jacksonville, FL, May 22–23, 2007. Samoilenko, S. and Osei-Bryson, K.M. (2007b). Increasing the Discriminatory Power of DEA in the Presence of the Sample Heterogeneity with Cluster Analysis and Decision Trees. Expert Systems with Applications, 34(2), 1568–1581. Samoilenko, S. (2008a). Information Systems Fitness and Risk in IS Development: Insights and Implications from Chaos and Complex Systems Theories. Information Systems Frontiers, 10(3), 281–292. Samoilenko, S. (2016). Disparity of Social and Economic Impact of ICT Capabilities in SubSaharan Economies: Empirical Investigation of Differentiating Factors. In Proceedings of the SIG GlobDev Pre-ECIS Workshop ICT in Global Development, Istanbul, Turkey, June 12, 2016. Samoilenko, S. and Osei-Bryson, K.M. (2008). An Exploration of the Effects of the Interaction between ICT and Labor Force on Economic Growth in Transitional Economies. International Journal of Production Economics, 115, 471–481. Samoilenko, S. and Osei-Bryson, K.-M. (2013). Using Data Envelopment Analysis (DEA) for Monitoring Efficiency-Based Performance of Productivity-Driven Organizations: Design and Implementation of a Decision Support System. Omega, 41(1), 131–142. Samoilenko, S. and Osei-Bryson, K.-M. (2017). A Methodology for Identifying Sources of Disparities in the Socio-Economic Impacts of ICT Capabilities in Sub-Saharan Economies. In Proceedings of the 2017 International Conference on Information Resources Management (conf-IRM 2017). http://aisel.aisnet.org/confirm2017/2 Techinsights.com. (2018). Apple IPhone 8 Plus Teardown. TechInsights, techinsights.com/ about-techinsights/overview/blog/apple-iphone-8-teardown/. Taylor, F. W. (1914). The Principles of Scientific Management. Harper. Worthington, A. C. and Dollery, B. E. (2002). Incorporating Contextual Information in Public Sector Efficiency Analyses: A Comparative Study of NSW Local Government. Applied Economics, 34(4), 453–464.

23

An Empirical Investigation of ICT Capabilities and the Cost of Business Start-up Procedures in Sub-Saharan African Economies

Investments in information and communication technologies (ICTs) are common to all economies of the world, and they are allocated to pursue common goals of positively impacting context-specific socio-economic targets. One of the underlying principles of the framework is that “NRI should provide clear policy guidance” (p. xi) to its users, which is not an easy undertaking to implement because within the framework “…the complex relationships between ICTs and socio-economic performance are not fully understood and their causality not fully established” (Di Battista, Dutta, Geiger, & Lanvin, 2015, p. 4). Simply put, Drivers of the framework are not directly associated with Impact. This is not surprising, for ICT Capabilities, reflected by Drivers, are simply opportunities that must be, in one form or another, taken advantage of in order to produce any socio-economic Impact. Perhaps, an illustration may help: an individual may have a set of capabilities qualifying her to work (e.g., level of education, level of experience, and level of social maturity), but this set of capabilities does not directly result in any level of socioeconomic impact (e.g., salary, title, level of seniority, and social status). Instead, in this case the proper chain of links representing “Capabilities → Impact” will be represented by the sequence “Capabilities → Employment → Impact”. Thus, capabilities, or drivers, must fuel some sort of an Engine that produces an Impact. Similarly, within the framework of NRI ICT capabilities reflected by Drivers must enable some sort of a socio-economic Engine that actually produces Impact. It is because of this absence of the direct association between Drivers and Impact, that the framework offers a great platform for exploring various economic means that power social and economic well-being of any economy. The overall framework for such exploration, consistent with the framework of NRI, is represented in Figure 23.1. It is worth noting that this framework is sufficient enough to investigate any type of Engine of the impact of ICT Capabilities, and it is minimal enough in allowing doing so in the most efficient manner. Furthermore, this framework is consistent with a grand economic theory – the framework of Neoclassical Growth Accounting (Solow, 1957). 183

184

Quantitative Methodologies using Multi-Methods

FIGURE 23.1  Framework for exploration of socio-economic impacts of ICT Capabilities.

According to the Neoclassical model, long-term economic growth is a product of capital investment, labor, and technological progress. To reconcile the framework of Figure 23.1 with the Neoclassical model, we only need to accept the fact that ICT Capabilities (e.g., drivers) of NRI are not given, but a result of appropriate investments in ICT. Furthermore, if we allow for a certain portion of socio-economic impact to be re-invested, then we will arrive at a framework of sustainable macroeconomic impact of investments in ICT similar to that of Samoilenko and OseiBryson (2017). The resultant framework is depicted in Figure 23.2. One of the recognized engines of socio-economic impact is Small and Medium Enterprises (SME), which is not only a source of over 50% of formally documented jobs in the world but also is a provider of effective solutions to critical development issues (TWB, 2016). The presence of SMEs is essential to the poorest countries in the world, those of Sub-Saharan Africa (SSA), where they serve as an important driver of economic growth that accounts for a majority of all businesses (IFC, 2017). Furthermore, African SMEs are also important to the global economy because their presences and success create “…a growing middle class with disposable income, in tandem with market opportunities for new investors” (de Sousa dos Santos, 2015)

FIGURE 23.2  Integrating framework of NRI with the framework of neoclassical growth accounting.

An Empirical Investigation of ICT Capabilities

185

and it is expected that over the next two decades Sub-Saharan economies “…will become the main source of new entrants into the global labor force” (IMF, 2015). It was noted, however, that while the SMEs of SSA are dominant players in their economies, they typically run unregistered businesses (Fierro, 2015). This has two important implications. First, informal enterprises do not participate in the formal economy, and, second, they have a very limited access to finance (Fjose, Grünfeld, & Green, 2010). Both implications are important, but it is an access to finance that was identified as a critical factor for the survival and growth of SSA’s SMEs (Fierro, 2015). There are various factors that may prevent informal enterprises from becoming registered SMEs, but we call the attention of our readers to an obvious one – Cost of Business Start-Up Procedures (CBSP) in a given economy. The justification for choosing this factor is an intuitive one, for CBSP must be incurred in order to formally register an SME. The importance of this factor is also straightforward, for formally registered SMEs have an easier access to finance required for sustaining and developing the business. In this chapter we investigate two broad questions within the context of SSA: 1. Whether ICT Capabilities impact the Cost of Business Start-Up Procedures (CBSP)? 2. Whether the cost of business start-up procedures is associated with a socio-economic impact? The conceptual underpinnings of these questions are simple. Let us consider the role of ICT and of the ICT Capabilities in any economy – their primary role is to create new information channels and/or to increase the efficiency and effectiveness of the existing ones. By fulfilling this role, ICT should reduce transaction costs and, as a result, decrease the CBSP. But does it do so effectively? Does it do so efficiently? We look closely at these questions in this chapter. There are at least two types of socio-economic impacts that CBSP can deliver. First one is associated with the consequences of the legalization of the business and all possible positive results of it. This type of an impact is a complex one, for it must account for a wide variety of benefits brought to all the beneficiaries of the legalization within the business (e.g., owners, employees, and other stakeholders) and within the context (e.g., social community, vendors, and customers) where the business operates. We consider the first type of impact to be outside of the scope of this chapter. The second type of impact is a much simpler one and is associated with a pure economic cost of legalization of the business. This cost is incurred by an individual, or individuals, and must be allocated out of the disposable income or a loan. It is precisely this type of impact that we investigate in this study. Let us now consider CBSP, in its pure and simple form of incurring associated fees, as an engine of socio-economic impact. First of all, it is only expected that a lower cost of legalization of SME produces direct economic impact, for if less money is allocated toward the start-up procedures, then more money will be available for business development or toward out-of-pocket spending. Similarly, we expect that the low CBSP will produce a positive social impact, for a registered business owner, ceteris paribus, enjoys a higher level of social recognition than an owner of unregistered business. Additionally, lower CBSP would leave more money to be allocated

186

Quantitative Methodologies using Multi-Methods

toward education and health care. But does, indeed, CBSP produce a socio-economic impact? We aim to answer this question in our inquiry. Essentially, we approach the subject of this inquiry from the perspective of a complex adaptive system, where the target system (e.g., a given SSA economy or group of economies) is comprised three components (e.g., ICT Capabilities, CBSP, and socioeconomic outcomes) bound by non-linear relationships. And it is, fundamentally, the context-dependent strength of the relationships is of primary concern to this study. At this point we can formulate the overall question that we aim to answer, as follows: Whether ICT Capabilities can result in a positive socio-economic outcome by impacting the CBSP in the context of SSA economies? We conduct our investigation within the context of 26 economies of SSA, using the data set for the period of 2012–2016. The analysis of the data is supported by a five-phase methodology utilizing cluster analysis (CA), data envelopment analysis (DEA), decision tree induction (DTI), association rule mining (ARM), and ordinary least squares (OLS).

THE RESEARCH FRAMEWORK AND RESEARCH QUESTIONS In our investigation we rely on the NRI framework (Kirkman, Osorio, & Sachs, 2002), the adapted version of which is depicted below in Figure 23.3. The framework relies on four sub-indexes and their ten sub-categories (or pillars) to obtain the value of the NRI, which reflects the capacity of economies to benefit from ICT. This research framework allows for answering a set of important questions. We offer the formulation of the questions (Q#:), along with the rationale for each question (R#:), below. Q1: Do SSA economies in our sample represent a homogenous group in regard to NRI and CBSP? R1: We know that SSA in the sample belongs to three different income groups; however, this differentiation is a received one (e.g., based on classification of World Bank), and is not based on our study’ factors (e.g., NRI and CBSP).

FIGURE 23.3  The research framework of the study.

An Empirical Investigation of ICT Capabilities

187

Q2: If the sample of SSA economies comprises multiple sub-groups, then what differentiates the heterogeneous sub-groups? R2: Knowing what differentiates heterogeneous sub-groups allows for additional information to be used in decision making; specifically, we can identify the sources of heterogeneity in the sample. Q3: In the case of the sample heterogeneity, how do sub-groups of SSA economies differ in terms of the relative efficiency of the conversion of ICT Capabilities into CBSP? R3: Knowledge of the relative efficiency of the conversion of ICT Capabilities into CBSP for each of the groups allows us to identify the best performers among the groups of SSA economies. This information is not only relevant to the research in the area of ICT4D, but it is also invaluable for the practical purposes of benchmarking and sharing the best practices in the context of SSA. Q4: Is there a set of actionable “If → Then” rules that could be generated to differentiate the best performers from the other groups in the sample? R4: By discovering a set of actionable “If → Then” rules that include common factors laggards will be able to benchmark the best performers in the sample. Furthermore, discovering naturally occurring causal structures allows for obtaining deeper insights regarding the mechanisms of socio-economic impact of ICT. Q5: Is there a relationship between the value of CBSP and socio-economic impact? R5: By answering this question we will know whether the groups of SSA economies differ in regard to the impact of CBSP.

PROPOSED METHODOLOGY The aim of this part of the chapter is to describe the purpose and the expected outcome of each of the five phases of the methodology, while leaving the technical details of each method out. The methodology of this investigation is an extension of the methodology of Samoilenko and Osei-Bryson (2016).

Phase 1: Cluster Analysis (CA) The purpose of CA is to identify naturally occurring sub-sets, or groupings, that may exist in the data set. While in our case SSA economies are categorized by the level of income, that categorization is based on the external criteria and is not CBSP- or NRIspecific. By performing CA, we can determine if SSA economies are alike in terms of NRI- and CBSP-specific criteria or if the sample is heterogeneous and consists of multiple sub-groups. The outcome of Phase 1 is N-cluster solution. It is worth noting that partitioning of the sample into sub-groups is logical in nature (Samoilenko and Osei-Bryson, 2008), where no physical division of the sample takes place.

188

Quantitative Methodologies using Multi-Methods

Phase 2: Decision Tree Induction If the results of Phase 1 yield a multi-cluster solution, then we create a variable target, with the domain of values being the number of identified clusters (e.g., 3-cluster solution will yield target {1,2,3}). By performing DTI, we will be able to identify, based on the top-level splits, the variables that differentiate the clusters the most. The outcome of Phase 2 is a set of variables differentiating N-clusters comprising the sample.

Phase 3: Data Envelopment Analysis During the third phase, we rely on DEA to evaluate the scores of relative efficiency of the “Drivers → CBSP” paths for each of N clusters in the sample. The purpose of using DEA in this investigation is to evaluate the level of efficiency with which ICT Capabilities transformed into CBSP. The outcome of Phase 3 is a set of averaged scores for each of N clusters, which would allow us to identify the benchmark group in the sample of SSA economies. We also conduct DEA to assess the relative efficiency of each of the clusters in terms of conversion of drivers into impact. We will use two models- the first one to assess the overall impact and the second to assess social and economic impact.

Phase 4: Ordinary Least Squares Regression The purpose of Phase 4 is to evaluate the presence of the causal relationships between CBSP and socio-economic outcomes. For each of N clusters of the sample, we will run OLS three times with the purpose of regressing CBSP against social, economic, and socio-economic impacts of NRI. The outcome of Phase 4 is a set (3 × N clusters) of p-values allowing for the testing of the significance of the relationships.

Phase 5: Association Rule Mining The purpose of Phase 5 is to find possible patterns, associations, or causal structures that may exist in our data. One of the main advantages of ARM is that it is suitable for undirected data mining; thus, we’ll aim to discover naturally occurring associations between the pillars of NRI and CBSP that may describe, in “If → Then” fashion, sub-groups of our sample. The outcome of Phase 5 is a set of actionable “If → Then” rules incorporating the variables discovered in Phase 2.

DATA We obtained the data from two publicly available and highly reputable sources. NRI’ data was downloaded from World Economic Forum’ Networked Readiness Index’ page (WEF_NRI, 2016) and the values of CBSP (% of GNI per capita) were obtained from the Data Bank of the World Bank (The World Bank, 2016). In 2012 the number and representation of the pillars of three sub-indexes, the drivers, of NRI has changed, and it was also the year when the impact sub-index was

An Empirical Investigation of ICT Capabilities

189

TABLE 23.1 Sample of Sub-Saharan Economies, by Income Level Income Level Low Low- Middle Upper- Middle

SSA Economies Burundi, Chad, Ethiopia, Gambia, Kenya, Madagascar, Malawi, Mali, Mozambique, Rwanda, Tanzania, Uganda, Zimbabwe Cameroon, Cape Verde, Côte d’Ivoire, Ghana, Lesotho, Nigeria, Senegal, Swaziland, Zambia Botswana, Mauritius, Namibia, South Africa

introduced. Given the changes that took place between 2011 and 2012, we decided to concentrate on the new version of NRI and collect the data for the period between 2012 and 2016. The inclusion of SSA economies in the WEF’ Global Information Technology Report (the source of the annual NRI data) is not always consistent. Resultantly, we could not include some of the SSA economies (e.g., Angola, Burkina Faso, Seychelles, Liberia, Gabon, Sierra Leone, and Guinea) in our sample simply because the data for some of the years were missing. Overall, we were able to compile the data set representing 26 economies of SubSaharan Africa (the classification of the International Monetary Fund as of October 2014). The sample consists of 13 low-income economies, nine low-middle economies, and four upper-middle economies (the classification of the World Bank as of July 2014). Membership of each group of the sample is provided in Table 23.1.

RESULTS OF THE DATA ANALYSIS Phase 1: CA While performing CA we have followed a rule of thumb for the relative size of any single cluster being greater than 10% of the overall size of the sample (Samoilenko and Osei-Bryson, 2008). By performing hybrid hierarchical clustering (there seems to be a consensus that this approach combines the advantages of other methods (Laan & Pollard, 2002; Chipman & Tibshirani, 2006) without having any unique disadvantages), we arrived at a 3-cluster solution. The results are summarized in Table 23.2. Based on the original solution, 19 SSA economies are “permanent residents” of a given cluster, while 7 SSA economies are “migrants”. This indicates that 77% of the clusters are stable in regard to their membership. Furthermore, if we consider four out of five years to be the indicator of stability, then over 92% of the clusters’ membership is stable.

Phase 2: DTI By performing DTI, we were able to identify sub-indexes and pillars that differentiate the three clusters of SSA economies in our sample. During this phase we constructed various decision trees, based on different variables (e.g., sub-indexes,

190

Quantitative Methodologies using Multi-Methods

TABLE 23.2 Results of Clustering: Membership, Income Composition, and Size of Each Cluster Cluster # Cluster 1: Size:32, A_CBSP = 29.05 Cluster 2: Size: 62, A_CBSP = 57.90

Cluster 3: Size: 36, A_CBSP = 49.43

Composition LI 21.88% LM 43.75% UM 34.38% LI 69.35% LM 30.65% UM 0.00% LI 27.78% LM 47.22% UM 25.00%

Cluster Membership Mauritius, South Africa, Cape Verde, Ghana, Zimbabwe (2013–2015), Botswana (2012), Uganda (2012–2015), Kenya (2014–2016) Burundi, Chad, Ethiopia, Madagascar, Malawi, Mali, Mozambique, Tanzania, Swaziland, Côte d’Ivoire (2012–2014), Nigeria (2015–2016), Uganda (2016), Lesotho (2012–2015), Zimbabwe (2012, 2016) Cameroon, Senegal, Zambia, Gambia, Rwanda, Namibia, Kenya (2012–2013), Côte d’Ivoire (2015–2016), Nigeria (2012–2013), Botswana (2013–2016)

pillars, NRI, and CBSP, level of income). We were able to determine, for example, that the three clusters cannot be clearly differentiated based on the income level, or based on the value of NRI, or based on impact sub-index of NRI. However, we obtained evidence that the clusters do differ based on the values of drivers (subindexes), and the values of the associated pillars. We present summary of the most relevant results of DTI in Table 23.3. TABLE 23.3 Results of DTI: English Rules, Split Variables, and Split Criteria English Rules Pillars-based DT B04 > 3.880 | B05 > 2.815: One {three=0, one=32 (100%), two=0} | B05 ≤ 2.815: two {three=1, one=0, two=10} B04 ≤ 3.880 | C07 > 3.220 | | A01 > 3.005: three {three=35 (97%), one=0, two=5} | | A01 ≤ 3.005: two {three=0, one=0, two=4} | C07 ≤ 3.220: two {three=0, one=0, two=43 (69.35%} Sub-index-based DT B > 3.725: One {three=0, one=29 (90.63%), two=0} B ≤ 3.725 | C > 2.950 | | A > 3.530: three {three=34 (94.44%), one=2, two=1} | | A ≤ 3.530: two {three=0, one=1, two=5} | C ≤ 2.950: two {one=0, three=2, two=56 (90.32%)}

Split Variables B04: Readiness sub-index: Affordability B05: Readiness sub-index: Skills A01: Environment Sub-index: Political and regulatory environment C07: Usage sub-index: Business usage

A- Environment sub-index B- Readiness sub-index C- Usage sub-index

An Empirical Investigation of ICT Capabilities

191

Phase 3: DEA During the third phase, we rely on DEA to evaluate the relative efficiency of “Drivers → CBSP” path for each of the three clusters. We will use variable return to scale (VRS) DEA model to conduct the analysis, for it is reasonable to argue that SSA economies have not yet reached the point of developing a level of ICT infrastructure allowing for accruing the benefits yielded by capitalizing on economies of scale. We also conduct DEA to evaluate the relative efficiency of “Drivers → Overall Impact” and “Drivers → (Social Impact, Economic Impact)” paths (Table 23.4). One point worth noting is the way of handling the variable Cost of Business Startup Procedures. DEA allows us to use three common models: input-, output-, and baseoriented. Input orientation deals with minimizing the level of inputs to achieve a given level of output. Output orientation is concerned with maximizing the level of outputs based on the given level of inputs. Base orientation refers to having control over inputs and outputs at the same time. In our case, we use an output-oriented model, for we are interested in the output-side of the DEA model. However, we are interested in minimizing CBSP, not in maximizing it. To deal with this situation we used a simple conversion scheme – first, we rounded up the greatest value of CBSP in the sample and labeled it CBSP Max. Then, for each data point, we subtracted the actual value of CBSP from CBSP Max and used that value in DEA. Results of the analysis presented in Table 23.5. It is interesting to note that while there is practically no difference between the clusters in regard to two Drivers → Impact models, there is a major difference between cluster 2 and the other two clusters in terms of Drivers → CBSP and two CBSP → Impact models.

Phase 4: OLS In this phase of our methodology, we run a regression analysis for each of the three clusters of SSA economies in our sample. For each cluster, we create three OLS TABLE 23.4 DEA Model of the Study DEA Model Drivers → CBSP

CBSP → Overall Impact CBSP → (Social Impact, Economic Impact) Drivers → Overall Impact

Drivers → (Social Impact, Economic Impact)

Inputs of DEA Model Environment Sub-index Readiness Sub-index Usage Sub-index CBSP CBSP Environment Sub-index Readiness Sub-index Usage Sub-index Environment Sub-index Readiness Sub-index Usage Sub-index

Outputs of DEA Model Cost of Business Startup Procedures Impact Sub-index Social Impact Pillar Economic Impact Pillar Impact Sub-index

Social Impact Pillar Economic Impact Pillar

192

Quantitative Methodologies using Multi-Methods

TABLE 23.5 Results of DEA

DEA Model Drivers → CBSP CBSP → Overall Impact CBSP → (Social Impact, Economic Impact) Drivers → Overall Impact Drivers → (Social Impact, Economic Impact)

Model Orientation Output-Oriented Input-Oriented Input-Oriented Input-Oriented Input-Oriented

Relative Efficiency Scores, Averaged Cluster 1 1.24 0.27 0.31 0.93 0.96

Cluster 2 4.49 0.006 0.006 0.94 0.96

Cluster 3 1.58 0.05 0.19 0.97 0.97

model, regressing the original value of CBPS against (1) socio-economic impact (sub-index), (2) social impact (pillar of impact sub-index), and (3) economic impact (pillar of impact sub-index). The results of OLS are summarized in Table 23.6.

Phase 5: ARM Overall, we ran ARM analysis four times. The first time we analyzed the complete sample (all three clusters) for the presence of general “If → Then” rules that would identify conditions for low-level CBSP. The results are provided in Table 23.7. We would like to remind our readers that the results of CA and DTI identified the following four variables impacting the heterogeneity of the sample: • • • •

Affordability Readiness pillar (Readiness sub-index) Skills Readiness pillar (Readiness sub-index) Political and regulatory Environment pillar (Environment sub-index) Business Usage pillar (Usage sub-index).

TABLE 23.6 Results of OLS Cluster 1

2

3

Depend. Var. Socio-economic Impact Social Impact Economic Impact Socio-economic Impact Social Impact Economic Impact Socio-economic Impact Social Impact Economic Impact

Adj. R 0.351 0.209 0.339 −0.014 −0.006 −0.016 0.120 0.045 0.139

F-value 17.772 9.172 16.864 0.150 0.617 0.025 5.790 2.652 6.658

P-value (95%) 0.000 0.005 0.000 0.700 0.435 0.875 0.022 0.113 0.014

Significant? Yes Yes Yes No No No Yes No Yes

193

An Empirical Investigation of ICT Capabilities

TABLE 23.7 Results of ARM, Complete Data Set Left Side (If) HIGH Political & Regulatory Environment HIGH Infrastructure Readiness HIGH Political & Regulatory Environment HIGH Individual Usage HIGH Political & Regulatory Environment HIGH Skills Readiness HIGH Political & Regulatory Environment HIGH Infrastructure Readiness HIGH Individual Usage HIGH Infrastructure Readiness HIGH Skills Readiness HIGH Individual Usage

& &

→ →

Right Side (Then) LOW CBSP

Sup. 0.14

Conf. 0.86

Lift 3.38

&



LOW CBSP

0.14

0.86

3.38

&



LOW CBSP

0.11

0.93

3.68

&



LOW CBSP

0.13

1.00

3.94

&



LOW CBSP

0.12

0.88

3.48

Consequently, we can identify the association rules that contain the variables (in bold) listed above. Results provided in Table 23.7 suggest that the level of CBSP is associated with the levels of two pillars – Political & Regulatory Environment and Skills Readiness. Because in this case ARM analysis was conducted using the complete data set, we have a basis to generalize that such association is may be valid for all SSA economies in our sample. Next, we ran ARM analysis three times – once for each of the sub-groups in the sample. By selecting the rules that contain variables that differentiate the subgroups, we were able to identify level-dependent associations that offer important actionable insights. We offer the summary of the results in Table 23.8. TABLE 23.8 Results of ARM, Summarized Comparisons of Three Clusters Left Side (If) HIGH Political & Regulatory Environment MIDLOW Political & Regulatory Environment HIGH Political & Regulatory Environment HIGH Individual Usage MIDLOW Political & Regulatory Environment MIDHIGH Individual Usage HIGH Skills Readiness HIGH Business Usage

→ →

Right Side (Then) LOW CBSP

Sup. 0.34

Conf. 0.85

Lift 1.69



MIDHIGH CBSP

0.19

0.88

2.10

&



LOW CBSP

0.31

0.83

1.67

&



MIDHIGH CBSP

0.19

0.88

2.10

&



HIGH EconImpact

0.31

0.91

2.08

&

(Continued)

194

Quantitative Methodologies using Multi-Methods

TABLE 23.8 (Continued) Left Side (If) LOW Business Usage LOW Government Usage HIGH Business Usage LOW Political & Regulatory Environment LOW Business Usage HIGH Business Usage HIGH Government Usage LOW Business Usage LOW Government Usage HIGH Political & Regulatory Environment HIGH Government Usage LOW Political & Regulatory Environment LOW Business Usage HIGH Business Usage LOW Business Usage LOW Government Usage

& &

→ →

Right Side (Then) LOW EconImpact

Sup. 0.29

Conf. 1.00

Lift 1.94

&

→ →

HIGH EconImpact LOW EconImpact

0.34 0.23

0.92 1.00

2.10 1.94

&



HIGH SocImpact

0.25

0.90

1.80

&



LOW SocImpact

0.27

0.94

1.95

&



HIGH SocImpact

0.25

1.00

2.00

&



LOW SocImpact

0.23

1.00

2.07

&

→ →

HIGH SocImpact LOW SocImpact

0.34 0.29

0.92 1.00

2.10 1.94

INTERPRETATION OF THE RESULTS OF THE DATA ANALYSIS In this section of the chapter, we offer to our readers a discussion of the results of our investigation. We proceed in the order of the steps of our methodology, where, for each step, we offer our interpretation of the results and briefly comment on the significance of the findings.

Cluster Analysis Despite sharing a common designation of SSA economies, the countries in our sample turned out to be quite different, and the difference was not along the lines of the received criterion (e.g., level of income) supplied by the World Bank. The three groups identified by the analysis turned out to be fairly stable in terms of their membership over a period of time. Interestingly, the membership of the sub-groups of the sample presented a mix in regard to the received categorization of the level of income (e.g., low income, low middle income, and upper middle income). There are three valuable insights that the results of CA provide to an investigator: 1. The presence of the commonly shared designation of SSA economies should not be mistaken for the indication of heterogeneity of SSA economies 2. The sources of heterogeneity of SSA economies are specific to the area of inquiry 3. Any received categorization (such as via income level) must be rigorously tested if the area of inquiry is different than that of received categorization.

An Empirical Investigation of ICT Capabilities

195

Decision Tree Induction The obtained information regarding the heterogeneity of our sample of SSA economies led us to inquiring into the sources of heterogeneity – this was accomplished by performing DTI analysis. Based on the framework of the study, there are three possible sources of heterogeneity-based on the value of NRI, based on the values of sub-indexes of NRI, and based on the pillars of sub-indexes of NRI. Additionally, we decided to explore if the heterogeneity of the clusters could be partially explained by the level of income. The obtained evidence suggested that neither NRI nor the level of income is among the variables differentiating the sub-groups. However, sub-indexes and pillars of sub-indexes are turned out to be effective at explaining the differences between the sub-groups. We decided to concentrate on pillars of the sub-indexes of NRI, for such choice offered a lower level of granularity for our inquiry and, as a result, a greater discriminatory power in explaining the differences between the sub-groups. The decision is easy to justify – we identified that only drivers’ sub-indexes and their respective pillars adequately explain the heterogeneity of the sample. Consequently, the choice was either to use three or eight variables – we chose the latter option. Overall, out of 10 pillars of NRI, there are four that explain the differences between the sub-groups of SSA economies – it was unexpected to find out that the values of 60% of the pillars of NRI play no role in differentiating SSA economies in our sample. Furthermore, it was surprising to find out that the countries in our sample do not discernably differ in terms of social and economic impacts of ICT capabilities, and that the heterogeneity of the sample could be adequately explained in terms of only four variables – Affordability Readiness, Skills Readiness, Political & Regulatory Environment, and Business Usage pillars of NRI. These findings offered some important insights, as follows: 1. The economies of SSA do differ in a fairly narrow way – a sufficient explanation for the presence of heterogeneity in the sample can be offered on the basis of only four pillars (40% of the available variables) of the framework of NRI 2. The sub-groups of SSA economies do not differ significantly on the basis of socio-economic impacts of ICT Capabilities (the output side of “Input → Output” equation) 3. The sub-groups of SSA economies do differ on the basis of ICT Capabilities (the input side of “Input → Output” equation) 4. Based on general “Input → Output” model, SSA economies do not differ in terms of efficiency and effectiveness of production of outputs (socioeconomic outcomes), but they do differ in terms of efficiency of utilization of inputs (ICT capabilities) 5. Overall, the insights suggest that the groups of SSA economies differ in terms of effectiveness and efficiency of transforming investments in ICT into a narrow sub-set (four pillars of NRI) of ICT capabilities (drivers of NRI).

196

Quantitative Methodologies using Multi-Methods

DEA We arrived to the third step of our methodology knowing that the sub-groups comprising the sample of SSA economies do differ in terms of ICT Capabilities (e.g., drivers of NRI), and that they do not differ in terms of socio-economic outputs (e.g., impact of NRI). We also determined that the three sub-groups do differ in regard to the values of CBSP – the Engine of transformation of Drivers into Impact. DEA evaluates the relative efficiencies of decision making units (DMU) – SSA economies – regarding conversion of inputs into outputs. By using five different DEA models as a test bed, we were able to obtain some valuable insights regarding the appropriate context models. Perhaps, it is worthwhile to comment on what in our view constitutes an appropriate DEA model. Let us consider the fact that scores of relative efficiency of the DMUs calculated via DEA are derived from a ratio of inputs to outputs specified by the DEA model. Consequently, given the differences in initial conditions – the levels of values of inputs and outputs, it is only expected that the differences should be reflected by the scores of DEA. Meaning, a DEA model should be able to translate the differences in initial conditions (e.g., values of inputs and outputs) into the differences in the scores of relative efficiency that are based on initial conditions. However, the results of DEA indicated that the three sub-group that do differ in regard to ICT Capabilities do not discernably differ in terms of the scores of relative efficiency of conversion of ICT Capabilities (e.g., Drivers) into socio-economic outcomes (e.g., impact). Resultantly, we consider “Drivers → Impact” model to be not appropriate for the context of our inquiry. At the same time, we considered two other DEA models to be appropriate – the models are “Drivers → Engine” (e.g., ICT Capabilities → CBSP) and “Engine → Impact” (e.g., CBSP → Socio-Economic Outcomes). The first model, ICT Capabilities → CBSP, accurately reflected the differences in levels of inputs and outputs and the impact of the differences on the overall scores of relative efficiency. Insights provided by the results of the analysis suggest that SSA economies belonging to Cluster 2 should significantly lower the level of output – values of CBSP. Similarly, the second model, CBSP → Socio-Economic Outcomes, suggests that the members of Cluster 2 should or, lower the levels of CBSP, or increase the values of socio-economic outcomes. Given the fact that the values of CBSP are driven by policies, and not economic necessities, the wise policy decision advice to SSA economies – members of Cluster 2 is to significantly lower the costs associated with the business start-up procedures. Overall, we summarize the importance of the findings provided by DEA as follows: 1. If using DEA for the purposed of investigating a relative efficiency of the conversion of ICT Capabilities into Socio-Economic Outcomes two DEA models of “ICT Capabilities → Engine of transformation” and “Engine of Transformation → Socio-Economic Outcomes” are preferred to a single model of “ICT Capabilities → Socio-Economic Outcomes” 2. Overall, the relative efficiency of conversion of ICT Capabilities into Socio-Economic Outcomes is impacted by the levels of CBSP that serves as an engine of conversion of capabilities into outcomes.

An Empirical Investigation of ICT Capabilities

197

Ordinary Least Squares (OLS) After we performed CA, we calculated the average value of the CBSP for each of the three sub-groups of our sample. It turned out that Cluster 1 has the lowest; Cluster 2 has the highest, while Cluster 3 being in between but closer to Cluster 2 in regard to the average value of CBSP. The purpose of performing OLS analysis was to find out whether the different levels of values of CBSP impact socio-economic outcomes of ICT Capabilities differently. Simply put, we wanted to find out if lower levels of CBSP are better for SSA economies than the higher levels. It worth pointing out that any value of CBSP that is greater than zero has a positive economic impact, for when start-up applicants pay the fee that money goes towards the economic bottom line. So, the question that we wanted to investigate is whether it is more beneficial to have an immediate and assured economic impact (e.g., via charging high fee), or if it is better to have a “delayed gratification” of charging less at the beginning and getting a greater socio-economic impact later when the money that could have been charged make their way to socio-economic outcomes. As we expected, the results of OLS were intuitive – lower values of CBSP were positively impacting social, economic, and the socio-economic impacts of ICT Capabilities, while high values of CBSP have no discernable impact. The obtained insight is important, especially for policy-making purposes, for 1. Higher levels of CBSP do not produce a discernable socio-economic impact, while the socio-economic impact of lower levels of CBSP is significant 2. One of the impacts of ICT Capabilities should be directed toward lowering of CBSP.

ARM By the time we performed ARM analysis, we knew that our sample of SSA economics is comprised three sub-groups, and we also knew the factors differentiating the sub-groups. As a result, we were able to utilize ARM to identify two important sets of rules. The first set comprises general rules that were generated on the basis of the complete sample of SSA economies. This set allows us to identify the general patterns that seem to hold within our context. The second set contains cluster-specific rules that are based on the dimensions (discovered via DTI) differentiating the sub-groups. This set offers an opportunity to not only examine cluster-specific associations but also to compare the associations between the clusters. Having such two sets of rules is important for three reasons. First, it allows for having two different perspectives – the global of the complete set and the local of each cluster. Second, it allows for merging of the two perspectives, where we can relate cluster-level rules to complete set-level rules. Third, it gives us an opportunity for meaningful local inter-cluster rule analysis while relating the analysis to the global set of rules.

198

Quantitative Methodologies using Multi-Methods

Overall, we interpret the obtained via ARM analysis insights as follows: 1. There is evidence that in the context of SSA economies high levels of Political & Regulatory Environment, Infrastructure Readiness, Skill Readiness, and Individual usage are associated with low levels of CBSP. 2. Consequently, if an SSA economy aims at having a low level of CBSP, then it is reasonable to suggest that investments in ICT should be directed at developing high-level capabilities in the areas of Political & Regulatory Environment, Infrastructure Readiness, Skill Readiness, and Individual Usage. 3. The results of ARM suggest that there is an inversely proportional relationship between the levels of Political & Regulatory Environment and CBSP. 4. There is evidence that the level of the economic impact of ICT Capabilities is directly proportional to the levels of Political & Regulatory Environment, Skill Readiness, and Business Usage. 5. There is also evidence that the level of the social impact of ICT Capabilities is directly proportional to the levels of Political & Regulatory Environment, Government Usage, and Business Usage.

DISCUSSION OF THE RESULTS OF THE STUDY Prior to conducting inquiry, we identified five research questions that served as a set of stepping stones leading to answering two main questions of the study. At this point, we are well equipped to provide the answers to each of the research questions, as well as to offer some “lesson learned” type of insights. Q1: Do SSA economies in our sample represent a homogenous group in regard to NRI and CBSP? A1: SSE economies do represent a homogeneous group in regard to the values of NRI, but they do differ in regard to the values of CBSP The obtained insight is intuitive – one should not assume the presence of homogeneity in the sample that is subjected to multi-dimensional analysis. Q2: If the sample of SSA economies is comprised heterogeneous sub-groups, then what differentiates the sub-groups? A2: Sub-groups of SSA economies differ in regard to the values of four pillars of NRI: Affordability Readiness, Skills Readiness, Political & Regulatory Environment, and Business Usage The obtained in this regard insight is related to dimensionality of the heterogeneity of the sample, where despite the presence of multiple dimensions potentially relevant to the analysis, it is usually a small sub-set of variables that will be identified (or selected) as the pertinent one to explain heterogeneity of the sample.

An Empirical Investigation of ICT Capabilities

199

Q3: In the case of the sample heterogeneity, how do sub-groups of SSA economies differ in terms of the relative efficiency of the conversion of ICT Capabilities into CBSP? A3: The sub-groups of SSA economies do differ in terms of the relative efficiency of the conversion of ICT Capabilities into CBSP, where the group with the high levels of ICT Capabilities and the low levels of CBSP is relatively more efficient than the group with the low level of ICT Capabilities and the high levels of CBSP It seems to be reasonable to suggest that if a DEA model contains variables that are also identified as sources of heterogeneity of the sample, then the heterogeneity of the sample will manifest in the results of DEA, thus yielding a sub-set of relatively efficient DMUs and a sub-set of relatively inefficient ones. Q4: Is there a set of actionable “If → Then” rules that could be generated to differentiate the best performers from the other groups in the sample? A4: There is a set of “If → Then” rules that differentiate the best performers from the rest of the group- in general, the following summary rule seems to hold true: “(HIGH Political & Regulatory Environment, HIGH Skills Readiness, HIGH Business Usage) → (LOW CBSP, HIGH Economic Impact, HIGH Social Impact)” The insight obtained from answering this question is not surprising – given the fact that any DEA model is expressed in the form “inputs/outputs”, it is of benefit to describe the difference between more and less relatively efficient groups of DMUs in the form of “(sub-set of Inputs) → (sub-set of Outputs)”. Q5: Is there a relationship between the value of CBSP and socio-economic impact? A5: Yes, there is a relationship between the value of CBSP and socio-economic impact and evidence suggests that SSA economies with low levels of CBSP tend to be associated with the high socio-economic impact, while SSA economies with the high levels of CBSP show no discernable socio-economic impact. The insight offered by this finding is a fairly important one, for it indicates that the mere presence of the engine of an impact (e.g., CBSP) does not guarantee the actual impact. Instead, the effectiveness of the engine of an impact is context-dependent and cannot be taken for granted. We would like to remind our reader that the subject of this investigation, to put it simply, was an inquiry into what can be called “two impacts”. First, we wanted to find out whether ICT Capabilities impact the CBSP. Second, we wanted to investigate the presence of the impact of the cost of business start-up procedures on socio-economic outcomes in the context of SSA economies. The results of our investigation allow for

200

Quantitative Methodologies using Multi-Methods

answering two broad questions of this study that we formulated at the beginning of our inquiry. The first question was: 1. Whether ICT Capabilities impact CBSP? According to the collected evidence, we can answer this question as follows: There is evidence that high levels of ICT Capabilities in the areas of Affordability Readiness, Skills Readiness, Political & Regulatory Environment, and Business Usage are associated with low levels of CBSP. The second question was: 2. Whether CBSP is associated with a socio-economic impact? The results of the data analysis offer sufficient evidence to answer the second question, as follows: The obtained evidence suggests that the socio-economic impact of CBSP is level-dependent, where low levels of the cost are associated with the presence of the impact, while high levels of the costs demonstrate no such association. By answering these questions, we are in a good position to address the overall question of our investigation, which is stated below: Whether ICT Capabilities can result in a positive socio-economic outcome by impacting the cost of business start-up procedures in the context of SSA economies? While based on the results of the data analysis the best short answer that we can provide to this question is “it depends”, we can formulate a longer version of the answer, as follows: High levels of ICT capabilities, especially in the areas of Affordability Readiness, Skills Readiness, Political & Regulatory Environment, and Business Usage are associated with lower levels of CBSP, and, in turn, with a positive socio-economic outcome. We believe that this answer offers a valuable insight to policy and decision-makers, for it does link investments in ICT and the socio-economic outcomes of such investments by a transparent “Input → Engine → Output” process.

CONCLUSION There is a consensus that SSA economies represent a group of poorest economies in the world, and one of the challenges for the researchers and practitioners working in the area of ICT for Development (ICT4D) is to offer and implement interventions that will change the status quo. Fundamentally, the challenges are of two types: first,

An Empirical Investigation of ICT Capabilities

201

find effective paths by which investments are transformed into socio-economic outcomes, and, second, find the ways to optimize the paths- make the mechanisms of transformation of investments into outcomes more efficient. Many SSA economies struggle with the problem of having a large number of informal SMEs and a low level of wealth, where formal SMEs contribute less than 20% to gross GDP versus around 60% in high-income countries (Fjose, Grünfeld, & Green, 2010). The link between the number of formal SMEs and the level of wealth of an economy, ceteris paribus, is apparent. The presence of this link is important, for it addresses the first challenge of finding an effective path of transformation of investments into economic outcomes via “investments → formalization of SMEs → socio-economic outcomes of investments” sequence of steps. And once the link is established, the optimization of the mechanism of transformation can take place. Consequently, it seems reasonable to suggest that one of the ways of improving the macroeconomic bottom line of SSA economies is by increasing the share of formal SMEs. Our investigation dealt with one of the barriers to establishing a formal SME – the cost of business start-up procedures. We can claim, based on the results of the data analysis, that this study was a successful one – we were able to, first, identify the context-specific factors that impact legitimization of SMEs in the context of SSA economies, and, second, establish a link between the relative magnitude of the legitimization barrier (e.g., CBSP) and socio-economic impact. One of the contributions of this investigation is a theoretical one – we have outlined a research framework allowing for investigating the socio-economic impact of ICT Capabilities via various engines of the impact. The suggested framework was not created from scratch; instead, it is an extension of the framework of NRI, and it is consistent with the grand theory of the Neoclassical Growth Accounting. Another contribution is a methodological one – despite the fact that Samoilenko and Osei-Bryson (2017) suggested using CA instead of received categorization to identify heterogeneous sub-groups in a sample, to our knowledge this study is the first one to actually implement it as a part of a methodology in action. We also believe that our study has some value for the practitioners working in the area of ICT4D – the results of our investigation offer some actionable insights, specifically in the form of “If → Then” rules, which could be put to practice in order to improve the socio-economic well-being of SSA economies. We must acknowledge limitations of our inquiry. The first one is associated with the availability of the data – we wish we could include more SSA economies in our sample, but we were limited by what was offered by our data sources. The second limitation is a fairly narrow focus on a single factor- that of CBSP. While, undoubtedly, CBSP is important to consider, there are other possible factors that may serve as barriers to legitimization of SMEs. Nevertheless, we hope that insights offered by this investigation outweigh its limitation.

ACKNOWLEDGMENT Material in this chapter previously appeared in: Start a Business, Get A Credit, Make an Impact: Do ICTs Help? Empirical investigation in the Context of Sub-Saharan Economies. In CONF-IRM (p. 38).

202

Quantitative Methodologies using Multi-Methods

REFERENCES Chipman, H. and Tibshirani, R. (2006). Hybrid Hierarchical Clustering with Applications to Microarray Data. Biostatistics Advance Access, 7(2), 286–301. de Sousa dos Santos, J. F. (2015). Why SMEs are key to growth in Africa. World Economic Forum. Available on-line at: https://www.weforum.org/agenda/2015/08/ why-smes-are-key-to-growth-in-africa/. Di Battista, A., Dutta, S., Geiger, T., and Lanvin, B. (2015). The Networked Readiness Index 2015: Taking the Pulse of the ICT Revolution. Global Information Technology Report 2015, pp. 3–28. Fierro, A. M. (2015). What are the biggest challenges for Africa’s entrepreneurs? World Economic Forum. Available on-line at: https://www.weforum.org/agenda/2015/08/ what-are-the-biggest-challenges-for-africas-entrepreneurs/. Fjose, S., Grünfeld, L., and Green, C. (2010). SMEs and Growth in Sub-Saharan Africa: Identifying SME Roles and Obstacles to SME Growth, MENON-publication no. 14/2010. MENON Business Economics. Available on-line at: https://www.norfund.no/ getfile.php/133983/Bilder/Publications/SME%20and%20growth%20MENON%20.pdf. IFC. (2017). SME Initiatives. International Finance Corporation (World Bank Group). Available on-line at: http://www.ifc.org/wps/wcm/connect/REGION__EXT_Content/Regions/ Sub-Saharan+Africa/Advisory+Services/SustainableBusiness/SME_Initiatives/. IMF. (2015). Regional Economic Outlook: Sub-Saharan Africa. World Economic and Financial Surveys of International Monetary Fund. Available on-line at: http://www. imf.org/external/pubs/ft/reo/2015/afr/eng/. Kirkman, S.G., Osorio, A.C., and Sachs, D.J. (2002). The Networked Readiness Index: Measuring the Preparedness of Nations for the Networked World. In The Global Information Technology Report 2001–2002 Readiness for the Networked World, Kirkman, G. (ed.). Oxford University Press, New York, NY, pp. 10–29. Laan, M. and Pollard, K. (2002). A New Algorithm for Hybrid Hierarchical Clustering with Visualization and the Bootstrap. Journal of Statistical Planning and Inference, 117 (2), 275–303. Samoilenko, S. and Osei-Bryson, K.M. (2008). Increasing the Discriminatory Power of DEA in the Presence of the Sample Heterogeneity with Cluster Analysis and Decision Trees. Expert Systems with Applications, 34(2), 1568–1581. Samoilenko, S. and Osei-Bryson, K.M. (2017). A Methodology for Identifying Sources of Disparities in the Socio-Economic Impacts of ICT Capabilities in Sub-Saharan Economies. In Proceedings of the International Conference on Information Resource Management, Santiago, Chile, May 16–19, 2017. Solow, R. (1957). Technical Change and the Aggregate Production Function. Review of Economics and Statistics, 39 (3), 312–320. The World Bank (2016). Data Bank: World Development Indicators. Available on-line at: http:// databank.worldbank.org/data/reports.aspx?source=world-development-indicators. TWB (2016). Entrepreneurs and Small Businesses Spur Economic Growth and Create Jobs. The World Bank. Available on-line at: http://www.worldbank.org/en/news/ feature/2016/06/20/entrepreneurs-and-small-businesses-spur-economic-growth-andcreate-jobs. WEF_NRI (2016). Networked Readiness Dataset. Available on-line at: http://reports.­ weforum.org/global-information-technology-report-2016/networked-readiness-index/.

24

Exploring the SocioEconomic Impacts of ICT-Enabled Public Value in Sub-Saharan Africa

INTRODUCTION Efficiency and effectiveness of a governmental apparatus are among the few qualities that are valued by citizens of any country. However, not every economy earns high marks in those areas (Kaufmann, Kraay, & Mastruzzi, 2010c); instead, most governments could use some improvement in the area of public services (Blaug et al., 2006). Consequently, interventions in this area (e.g., e-government initiatives) are common to most of the economies of the world. It was suggested that the practices and ideas of New Public Management (NPM) could serve as a tool for making administrative systems more reliable, consistent, and efficient (Hood & Lodge, 2006). One of the ways of implementing NPM is by adopting information and communication technologies (ICTs) to create ICT capabilities that reduce cost and time of organizational activities (Cordella, 2006). However, the actual impact of ICT in this regard is not easy to assess because the formulations of NPM are quite diverse (Cordella & Bonina, 2012). Additionally, it was also noted that the commonly identified impacts of ICT on public administration in terms of economics, effectiveness, and efficiency of the provided services represent too narrow of a focus (Cordella & Bonina, 2012); thus, social impacts should also be taken into consideration. Unsurprisingly, the relevant literature (Bannister & Connolly, 20141; Cordella & Willcocks, 2010) suggests that the notion of public value (Moore, 1995) is better suited to assess the impact of ICT on public initiatives (Cordella & Bonina, 2012). In this context public value can be defined as a benefit that is “…related to the achievements of objectives set by government programs and the delivery of public services to the citizens” (ibid, p. 516). It would appear then that the link “ICT Capabilities → Public Value” link is one that needs to be investigated. However, a Public Value created as a result of ICT-enabled public sector initiatives is not an end in itself, but a means to an end – it was noted that investigating socio-economic outcomes of ICT-enabled public initiatives is of benefit and is preferable to inquiries that are more limited in scope (Cordella & Bonina, 2012). Furthermore, the question seems to be is not if ICT Capabilities allow for generating Public Value (e.g., its effectiveness), but how efficiently such value is generated (Jaeger, 2005; Eppler, 2007). Consequently, given the three constructs – ICT 203

204

Quantitative Methodologies using Multi-Methods

Capabilities, Public Value, and socio-economic outcomes, we can state a general question of our inquiry, namely: Do ICT Capabilities deliver positive socio-economic outcomes via impacting Public Value creation? This question is worth investigating in general (WEF_GITR, 2016), but it gains an additional importance in the context of the poorest countries of the world that we chose to be the subject of our inquiry – the economies of Sub-Saharan Africa (SSA). Three obvious reasons for why this is so come to mind. Firstly, results of the previous research indicated that SSA could boost their economic development via expanding their public sector (Kimaro, Keong, & Sea, 2017), and the expansion of the public sector could be positively impacted by ICT capabilities. This will necessitate the investments in ICT, for ICT capabilities cannot be created without the appropriate investments. In sum, SSA economies will have to invest in ICT. Secondly, poorer economies have fewer resources to invest in ICT, and, as a result, they will be more dependent on efficiency of utilization of investments in ICT, as well as of efficiency of the impact of the resultant ICT Capabilities. Simply put, SSA economies are more vulnerable to the consequences of inefficiency of utilization of investments. Thirdly, it was noted that Public Value was found to influence the socio-economic development of SSA. There is evidence that bad governance negatively impacts economic performance (Habtamu, 2008); thus, it is only reasonable to expect that the reverse is also true and that ICT Capabilities-enabled “good governance” (e.g., generated Public Value) will have a positive impact on development of SSA economies. Conceptually, the causal structure of the general question that we stated can be presented in the form of “ICT Capabilities → Public Value → Socio-Economic Outcomes”. However, this simple deterministically causal chain of links is too simplistic for generating promising research questions – a more rigorous framework is needed. Unsurprisingly, the importance of developing theoretical underpinning linking ICT and public sector initiatives has been previously noted (Bekkers & Homburg, 2007; Madon, Sahay, & Sudan, 2007) and a possible mechanism linking the two (e.g., via efficient and effective information channels) has been acknowledged (Dunleavy et al., 2006; Gupta, Dasgupta, & Gupta, 2008). The first item on our agenda, therefore, is the placement of our inquiry on a solid theoretical footing that can support answering the general question. In the next part of the chapter we outline the research framework of this investigation to our readers.

RESEARCH FRAMEWORK OF THE STUDY In this investigation we rely on the framework of Networked Readiness Index (NRI), which serves as an established measure of ICT capabilities, or, drivers, of socioeconomic development (WEF_GITR, 2015). The NRI framework is reflected by a single composite index (e.g., NRI) comprised of four sub-indexes – three drivers (Environment, Readiness, and Usage sub-indexes) and one impact sub-index.

ICT-Enabled Public Value in SSA

205

FIGURE 24.1  The research framework of the study.

Drivers of the framework are not directly associated with impact. And it is because of this absence of the direct association between drivers and impact that the framework of NRI offers a great platform for exploring various factors (e.g., serving as an engine of the impact) that contribute to social and economic well-being of economies. For the purposes of our investigation we adapt the framework of NRI by using public value as the engine of the impact – the resultant framework is depicted in Figure 24.1. This framework allows us to state two broad questions that we aim to answer within the context of SSA: 1. How efficiently ICT Capabilities impact the creation of Public Value? 2. Whether the created Public Value is associated with a socio-economic impact? The conceptual underpinnings of these questions are simple. Let us consider the role of ICT and of the ICT Capabilities in any economy – their primary role is to create new information channels and/or to increase the efficiency and effectiveness of the existing ones. The role of ICT implementations in government and public services is no different in this regard (Eppler, 2007; Jaeger, 2005). By fulfilling this role ICT should reduce transaction costs and, as a result, improve the quality of public services and, consequently, generate public value. But does it do so effectively? Does it do so efficiently? We look closely at these questions in this study. To proceed further we need to clarify to our readers the type of an impact of Public Value that we aim to investigate. There are at least two types of socio-economic impacts that Public Value can deliver. First one is associated with the consequences of decisions made by the citizens that receive public value as a result of effective and efficient functioning of their government. This type of an impact is a complex one, for it must account for a wide variety of benefits brought to all the beneficiaries of the received Public Value within the context where the value is provided. Let us consider rule of law as an instance of

206

Quantitative Methodologies using Multi-Methods

Public Value that a government provides to its citizenry. In a general sense rule of law refers to the extent to which citizens have confidence in rules of society, and this perception of citizens will have a wide impact on their decisions regarding participation in any and all areas of life that involve social contracts. We consider the first type of the impact, decisional, to be outside of the scope of this chapter – it is simply too complex for us to handle. The second type of an impact is much simpler and is associated with a pure socioeconomic benefit of receiving public value at a point of performing a social transaction. Simply put, in order to perform any sort of social transaction in the area where public value is provided an individual must incur certain costs (e.g., time, money, energy). Thus, this type of impact is transactional in nature, and here we state a set of premises on which we rely in this investigation, as follows: 1. Public Value provides a transactional benefit to the citizenry by impacting a socio-economic cost of social transactions, where greater Public Value results in lower socio-economic costs of social transactions. 2. Delivery and creation of Public Value is associated with effectiveness and efficiency of the agency tasked with the provision of the value. 3. ICT Capabilities can positively impact efficiency and effectiveness of the agency delivering public value via creation of new and optimization of the existing information channels. It is precisely this, transactional, type of the impact that we investigate in this study. Let us now consider Public Value as an engine of socio-economic impact. First of all, it is only expected that a lower cost of performing a social transaction produces a direct economic impact, for if less money is allocated towards the transaction, then more money will be available for everything else, from business development to simple out-of-pocket spending. Similarly, we expect that a low social cost of transaction will produce a positive social impact, for less time and energy is allocated towards the transaction the more is left for other things and activities. But does, indeed, Public Value produce a socio-economic impact? We aim to answer this question in our inquiry. Essentially, we approach the subject of this inquiry from the perspective of complex adaptive system, where the target system (e.g., a given SSA economy or group of economies) is comprised of three components (e.g., ICT Capabilities, Public Value, socio-economic outcomes) bound by non-linear relationships. And it is, fundamentally, the context-dependent strength of the relationships is of primary concern to this study. At this point we can formulate the overall question that we aim to answer, as follows: Could ICT Capabilities result in a positive socio-economic outcome by impacting the creation of Public Value within the context of SSA economies? We conduct our investigation within the context of 26 economies of SSA, using the data set for the period of 2012–2016. The analysis of the data is supported by a five-phase methodology utilizing cluster analysis (CA), data envelopment analysis

ICT-Enabled Public Value in SSA

207

(DEA), decision tree induction (DTI), association rule mining (ARM), and ordinary least squares (OLS).

RESEARCH QUESTIONS OF THE STUDY In our investigation we rely on the NRI framework (Kirkman, Osorio, & Sachs, 2002), the adapted version of which is depicted in Figure 24.2. An increase in the value of NRI for a given economy is indicative of the increase of the impact of ICT on innovation and productivity (Dutta & Jain, 2003). In this investigation we represent Public Value via the dimensions provided by Worldwide Governance Indicators (WGI) of the World Bank (The World Bank, 2016). The dimensions are Political Stability and Absence of Violence/Terrorism, Voice and Accountability, Rule of Law, Regulatory Quality, Government Effectiveness, and Control of Corruption. It is only fair to our reader if we proactively deal with the elephant in the room that is an issue of representation of Public Value, and offer our answer to the question of “Why could the chosen dimensions serve as an adequate representation of Public Value?” It is worth noting that the issue of representation of any concept is a valid one to consider, for rarely there is a consensus regarding the choice of variables adequately describing a construct, let alone the scale of measurement of the variables. The issue is philosophical, as well as practical in nature, and it can never be completely settled due to the inherent limitation of modeling – the incompleteness of any representation. However, what about the accuracy of representation? Do six Worldwide Governance Indicators accurately represent governance and, by extension, Public Value? In this chapter we take a position that the accuracy of representation of any non-trivial construct is fundamentally subjective and a consensus-based for a given domain. This is also, and may be even especially true, regarding the representation and measurement of governance, where the inherent difficulties of the undertaking

FIGURE 24.2  The Unpacked Research Framework of this Study.

208

Quantitative Methodologies using Multi-Methods

have been duly noted by Kaufmann, Kraay, and Mastruzzi (2010a). However, while construct validity of WGI is not immune to challenges (e.g., notably, Thomas, 2010), in the absence of better alternatives the definition of governance as it is supplied via WGI is perfectly acceptable for cross-country and over-time comparisons (Kaufmann et al., 2010b). In summary, despite the marketplace for development of governance indicators being wide open (Kaufmann et al., 2007), WGI remains a long-standing, legitimate, and widely accepted tool for a representation of governance (Kaufmann, et al., 2010c). This may also be a good opportunity to present our reader with yet another proposition supporting this study, which links WGI, governance, and public value, as follows: WGI can be used to represent governance, and one of the functions of governance is provision of Public Value to its citizenry. The representation of Public Value via WGI yields the framework depicted in Figure 24.2. Our reader will notice that the only difference between frameworks in Figures 24.1 and 24.2 is due to “unpacking” of public value into six indicators. The research framework of this study allows for answering a set of important questions. We offer the formulation of the questions (Q#:) below. Q1: Do SSA economies in our sample represent a homogenous group in regard to NRI and Public Value? Q2: If the sample of SSA economies is comprised of multiple sub-groups, then what differentiates the heterogeneous sub-groups? Q3: In the case of the sample heterogeneity, how do sub-groups of SSA economies differ in terms of the relative efficiency of the conversion of ICT Capabilities into Public Value? Q4: Is there a set of actionable “If → Then” rules that could be generated to differentiate the best performers from the other groups in the sample? Q5: Is there a relationship between the estimates of Public Value and the level of socio-economic impact?

METHODOLOGY OF THE INVESTIGATION We identified the methodology of Samoilenko and Osei-Bryson (2017) as being suitable for the purposes of our study. Thus, we adopted the methodology with a minor modification of adding CA and OLS regression. For the sake of saving space, we do not provide the full description of the methodology, but refer our readers to the paper of Samoilenko and Osei-Bryson (2017). A very brief overview of the methodology is provided below.

ICT-Enabled Public Value in SSA

209

Phase 1: Cluster Analysis The purpose of CA is to identify naturally occurring sub-sets, or groupings, that may exist in the data set. The outcome of Phase 1 is N-cluster solution – partitioning of the sample into sub-groups is logical in nature (Samoilenko and Osei-Bryson, 2008), where no physical division of the sample takes place.

Phase 2: Decision Tree Induction DTI allows us to identify, based on the top-level splits, the variables that differentiate the clusters the most. The outcome of Phase 2 is a set of variables differentiating N-clusters comprising the sample.

Phase 3: Data Envelopment Analysis (DEA) The purpose of using DEA in this investigation is to evaluate the level efficiency with which ICT Capabilities transformed into public value. The outcome of Phase 3 is set of averaged scores for each of N clusters, which would allow us to identify the benchmark group in the sample of SSA economies.

Phase 4: Ordinary Least Squares (OLS) Regression The purpose of Phase 4 is to evaluate the presence of the causal relationships between public value and socio-economic outcomes. The outcome of Phase 4 is a set (21×N clusters) of p-values allowing for the testing of the significance of the relationships.

Phase 5: Association Rule Mining (ARM) ARM allows us to find possible patterns, associations, or causal structures that may exist in our data. The outcome of Phase 5 is a set of actionable “If → Then” rules based on the variables discovered in Phase 2.

DATA We obtained the data from two publicly available and highly reputable sources. NRI’ data was downloaded from World Economic Forum’ Networked Readiness Index’ page (WEF_NRI, 2016) and the values of the variables representing Public Value were obtained from the Data Bank of the World Bank, Worldwide Governance Indicators database (The World Bank, 2016). In 2012 the number and representation of the pillars of three sub-indexes, the drivers, of NRI has changed, and it was also the year when the impact sub-index was introduced. Given the changes that took place between 2011 and 2012, we decided to concentrate on the new version of NRI and collect the data for the period between 2012 and 2016. Overall, we were able

210

Quantitative Methodologies using Multi-Methods

TABLE 24.1 Sample of Sub-Saharan Economies, by Income Level Income Level Low

Sub-Saharan Economies Burundi, Chad, Ethiopia, Gambia, Kenya, Madagascar, Malawi, Mali, Mozambique, Rwanda, Tanzania, Uganda, Zimbabwe Cameroon, Cape Verde, Côte d’Ivoire, Ghana, Lesotho, Nigeria, Senegal, Swaziland, Zambia Botswana, Mauritius, Namibia, South Africa

Low-Middle Upper-Middle

to compile the data set representing 26 economies of SSA (the classification of the International Monetary Fund as of October 2014). Membership of each group of the sample is provided in Table 24.1.

RESULTS OF THE DATA ANALYSIS Phase 1: CA While performing CA we have followed a rule of thumb for the relative size of any single cluster being greater than 10% of the overall size of the sample (Samoilenko & Osei-Bryson, 2008). By performing hybrid hierarchical clustering (there seems to be a consensus that this approach combines the advantages of other methods (Laan & Pollard, 2002; Chipman & Tibshirani, 2006) without having any unique disadvantages), we arrived to 3-cluster solution. Results are summarized in Table 24.2. We also calculated relevant average values of WGI per cluster. Results are presented in Table 24.3. Based on the original solution 19 SSA economies are “permanent residents” of a given cluster, while 7 SSA economies are “migrants”. This indicates that 77% of TABLE 24.2 Results of Clustering: Membership, Income Composition, and Size of Each Cluster Cluster # Cluster 1 Size:32 Cluster 2 Size: 62

Cluster 3 Size: 36

Composition LI 21.88% LM 43.75% UM 34.38% LI 69.35% LM 30.65% UM 0.00% LI 27.78% LM 47.22% UM 25.00%

Cluster Membership Mauritius, South Africa, Cape Verde, Ghana, Zimbabwe (2013– 2015), Botswana (2012), Uganda (2012–2015), Kenya (2014–2016) Burundi, Chad, Ethiopia, Madagascar, Malawi, Mali, Mozambique, Tanzania, Swaziland, Côte d’Ivoire (2012–2014), Nigeria (2015–2016), Uganda (2016), Lesotho (2012–2015), Zimbabwe (2012, 2016) Cameroon, Senegal, Zambia, Gambia, Rwanda, Namibia, Kenya (2012–2013), Côte d’Ivoire (2015–2016), Nigeria (2012–2013), Botswana (2013–2016)

211

ICT-Enabled Public Value in SSA

TABLE 24.3 Comparison of Tree Clusters in Terms of the Values of WGI and Sub-Indexes of NRI Voice and Rule of Regulatory Government Control of Political Average Cluster # Accountability Law Quality Effectiveness Corruption Stability of WGI Cluster 1  0.28 0.04  0.01 −0.04 −0.03 −0.16 −0.03 Cluster 2 −0.70 −0.74 −0.73 −0.86 −0.79 −0.83 −0.77 Cluster 3 −0.31 −0.24 −0.17 −0.32 −0.11 −0.08 −0.20 Cluster 1 Cluster 2 Cluster 3

Environment 4.05 3.30 4.04

Readiness 4.15 2.69 3.15

Usage 3.36 2.64 3.23

Impact 3.23 2.66 3.28

the clusters are stable in regard to their membership. Furthermore, if we consider four out of five years to be the indicator of stability, then over 92% of the clusters’ membership is stable.

Phase 2: DTI By performing DTI, we were able to identify sub-indexes and pillars that differentiate the three clusters of SSA economies in our sample. During this phase we constructed various decision trees, based on different target variables (e.g., sub-indexes, pillars, NRI, and level of income). We were able to determine, for example, that the three clusters cannot be clearly differentiated based on the income level, or based on the value of NRI, or based on impact sub-index of NRI. However, we obtained evidence that the clusters do differ based on the values of drivers (sub-indexes), and the values of the associated pillars. We present summary of the most relevant results of DTI in Table 24.4. TABLE 24.4 Results of DTI: English Rules, Split Variables, and Split Criteria English Rules Pillars-based DT B04 > 3.880 | B05 > 2.815: One {three=0, one=32 (100%), two=0} | B05 ≤ 2.815: two {three=1, one=0, two=10} B04 ≤ 3.880 | C07 > 3.220 | | A01 > 3.005: three {three=35 (97%), one=0, two=5} | | A01 ≤ 3.005: two {three=0, one=0, two=4} | C07 ≤ 3.220: two {three=0, one=0, two=43 (69.35%}

Split Variables Pillars B04 – Readiness sub-index: Affordability B05 – Readiness sub-index: Skills A01 – Environment sub-index: Political and regulatory environment C07 – Usage sub-index: Business usage

212

Quantitative Methodologies using Multi-Methods

Phase 3: DEA During the third phase we rely on DEA to evaluate relative efficiency of various “Drivers → public value” paths for each of the three clusters. We will use input-oriented model and variable return to scale (VRS) DEA model to conduct the analysis, for it is reasonable to argue that SSA economies have not yet reached the point of developing a level of ICT infrastructure allowing for accruing the benefits yielded by capitalizing on economies of scale. Description of the specific DEA models that we will be using in this phase is provided in Table 24.5. Results of the DEA presented in Table 24.6. It is interesting to note that while there is practically no difference between the clusters in regard to two Drivers → Impact models, there is a significant difference between cluster 2 and the other two clusters in terms of Drivers → Public Value models. Simply put, cluster 2 is relatively less efficient in converting ICT capabilities into public value.

Phase 4: OLS In this phase of our methodology we run a regression analysis for each of the three clusters of SSA economies in our sample. For each cluster we create 21 OLS models, regressing the average of all six WGI’, and each WGI, against, (1) Socioeconomic impact (sub-index), (2) Social impact (pillar of impact sub-index), and (3) economic impact (pillar of impact sub-index). The results are summarized in Table 24.7.

TABLE 24.5 DEA Model of the Study DEA Model Drivers → Ave.Pub.Val. Drivers → TotalPub.Val Drivers → Voice & Accountability Drivers → Rule of Law Drivers → Regulatory Quality Drivers → Government Effectiveness Drivers → Control of Corruption Drivers → Political Stability Drivers → Overall Impact Drivers → Social + Economic Impact

Inputs of DEA Model Environment, Readiness, Usage Sub-indexes Environment, Readiness, Usage Sub-indexes Environment, Readiness, Usage Sub-indexes

Outputs of DEA Model Average of 6 values of WGI All six indicators of WGI Voice & Accountability

Environment, Readiness, Usage Sub-indexes Environment, Readiness, Usage Sub-indexes

Rule of Law Regulatory Quality

Environment, Readiness, Usage Sub-indexes

Government Effectiveness

Environment, Readiness, Usage Sub-indexes

Control of Corruption

Environment, Readiness, Usage Sub-indexes

Political Stability

Environment, Readiness, Usage Sub-indexes Environment, Readiness, Usage Sub-indexes

Impact Sub-index Social Impact Pillar Economic Impact Pillar

213

ICT-Enabled Public Value in SSA

TABLE 24.6 Results of DEA Relative Efficiency Scores, Averaged DEA Model Drivers → Public Value (average of WGI) Drivers → Public Value (all 6 of WGI) Drivers → Voice & Accountability Drivers → Rule of Law Drivers → Regulatory Quality Drivers → Government Effectiveness Drivers → Control of Corruption Drivers → Political Stability Drivers → Overall Impact Drivers → (Social Impact, Economic Impact)

Model Orientation Input-oriented

Cluster 1 0.96

Cluster 2 0.91

Cluster 3 0.97

Input-oriented

0.99

0.96

0.99

Input-oriented Input-oriented Input-oriented Input-oriented

0.95 0.96 0.96 0.98

0.90 0.91 0.90 0.92

0.97 0.97 0.98 0.98

Input-oriented Input-oriented Input-oriented Input-oriented

0.93 0.93 0.93 0.96

0.90 0.89 0.94 0.96

0.98 0.97 0.97 0.97

TABLE 24.7 Results of OLS OLS, Independent Variable Average WGI Voice & Accountability Rule of Law Regulatory Quality Government Effectiveness Control of Corruption Political Stability Overall significant (out of 6) Effectiveness (average)

Significant at 5% Level? OLS, Dependent Variable Economic Impact Social Impact Economic Impact Social Impact Economic Impact Social Impact Economic Impact Social Impact Economic Impact Social Impact Economic Impact Social Impact Economic Impact Social Impact Economic Impact Social Impact (Economic + Social Impact)/2

Cluster 1 Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes No No 83% 83% 83%

Cluster 1 No Yes Yes Yes No Yes No No No Yes No Yes No No 16% 67% 41%

Cluster 1 Yes No Yes Yes Yes No No No Yes No Yes No Yes No 83% 16% 50%

214

Quantitative Methodologies using Multi-Methods

Phase 5: ARM We used R to run ARM analysis (the threshold was set at 20% for support and 60% for confidence). Overall, we ran ARM analysis 19 times. The first time we analyzed the complete sample (all three clusters) for the presence of general “If → Then” rules that would involve any of the indicators of WGI on the right (“Then”) side. After that we ran ARM six times, for each cluster, to identify rules that are specific to each of the indicators of WGI. The summarized results are provided in Table 24.8, while a larger set of chosen rules is provided in Appendix of the chapter. Looking at the rules provided in Table 24.8 our reader can easily notice two things. First, there seem to be an inverse relationship between ICT capabilities (as it is represented by the pillars of NRI) and public value (as represented by WGI). Second, there is a significant number of the association rules that contain Political and Regulatory Environment and Business Usage on the left side of “If → Then” equation – these are the variables that differentiate SSA economies in our sample.

TABLE 24.8 Results of ARM, Summary Left Side (If) HIGH Political & Regulatory Envir HIGH Political & Regulatory Envir HIGH Skills Readiness

→ →

Right Side (Then) LOW Reg. Quality

Sup. 0.21

Conf. 0.82

Lift 2.98



LOW Rule of Law

0.20

0.76

2.98



0.20

0.75

2.98

HIGH Infrastructure Readiness



0.20

0.62

2.38

HIGH Political & Regulatory Envir LOW Political & Regulatory Envir LOW Skills Readiness LOW Political & Regulatory Envir LOW Political & Regulatory Envir Control of Corruption 1 HIGH Political & Regulatory Envir 2 LOW Political & Regulatory Envir 3 HIGH Political & Regulatory Envir HIGH Infrastructure Readiness



LOW Government Effect LOW Government Effect LOW Control of Corruption HIGH Rule of Law

0.20

0.67

2.63

0.20

0.76

2.98

HIGH Political Stability HIGH Government Effect HIGH Control of Corruption

0.20 0.20

0.6 0.75

2.09 2.98

0.20

0.64

2.51

MidLo Control of Corruption HIGH Control of Corruption LOW Control of Corruption

0.26

1.00

2.11

0.30

0.64

2.00

0.28

1.00

2.25

# Complete Set

→ → → →

→ → →

215

ICT-Enabled Public Value in SSA # Left Side (If) Government Effectiveness 1 HIGH Infrastructure Readiness



Right Side (Then)

Sup.

Conf.

Lift



0.47

0.83

1.40

2



LOW Government Effectiv HIGH Government Effectiv

0.21

1.0

2.39



LOW Government Effectiv

0.28

0.83

2.14



LOW Political Stability

0.25

0.8

1.97



HIGH Political Stability

0.21

0.72

1.91



LOW Political Stability

0.22

0.89

2.29



LOW Rule of Law

0.40

1.0

1.53



HIGH Rule of Law

0.21

1.00

2.39



LOW Rule of Law

0.25

0.90

2.70



LOW Regulatory Quality

0.37

1.01

1.88



HIGH Regulatory Quality LOW Regulatory Quality

0.21

0.93

2.38

0.22

0.89

2.00



LOW Voice & Account

0.40

1.0

1.53

→ →

HIGH Voice & Account LOW Voice & Account

0.22 0.22

0.93 0.89

2.00 2.67

3

LOW Political & Regulatory Envir LOW Business & Innovation Envir HIGH Infrastructure Readiness

Political Stability 1 HIGH Affordability Readiness HIGH Skills Readiness HIGH Government Usage 2 LOW Affordability Readiness LOW Business Usage 3 HIGH Political & Regulatory Envir HIGH Individual Usage Rule of Law 1 HIGH Political & Regulatory Envir 2 LOW Political & Regulatory Envir LOW Business & Innovation Envir 3 HIGH Political & Regulatory Envir HIGH Infrastructure Readiness Regulatory Quality 1 HIGH Political & Regulatory Envir HIGH Business & Innovation Envir 2 LOW Business Usage 3

HIGH Political & Regulatory Envir HIGH Individual Usage Voice & Accountability 1 HIGH Political & Regulatory Envir 2 LOW Business Usage 3 HIGH Political & Regulatory Envir HIGH Individual Usage



216

Quantitative Methodologies using Multi-Methods

DISCUSSION OF THE RESULTS OF THE STUDY Prior to conducting inquiry, we identified five research questions that served as a set of stepping stones leading to answering two main questions of the study. At this point we are well-equipped to provide the answers (A#) to each of the research questions, below. A1: SSE economies do not constitute a homogeneous group in regard to their levels of ICT capabilities. A2: The dimensions differentiating heterogeneous groups of SSA are represented by four pillars of NRI: Affordability Readiness, Skills Readiness, Political and Regulatory Environment, and Business Usage. A3: The heterogeneous groups of SSA economies also differ in terms of the relative efficiency of the impact of ICT capabilities on public value, where the groups with higher levels of ICT capabilities and higher levels of WGI are relatively more efficient than the group with a lower level of ICT capabilities and a lower level of WGI. A4: There is also a set of “If → Then” rules that differentiates the groups of SSA economies – in general high values of pillars of NRI are associated with low values of WGI, and vice versa. A5: The obtained evidence suggests that the socio-economic impact of Public Value is level-dependent, and SSA economies with greater values of WGI are more effective in obtaining a socio-economic impact then the economies with the lower levels of WGI. The results of the data analysis allow for answering two broad questions of this study that we formulated in the beginning of our inquiry; the answers are as follows: 1. There is evidence that high levels of ICT Capabilities in the areas of Affordability Readiness, Skills Readiness, Political and Regulatory Environment, and Business Usage allow for relatively more efficient generation of Public Value. 2. The obtained evidence suggests that the socio-economic impact of Public Value is level-dependent, where economies with higher levels of WGI are more effective in producing the impact. By answering these questions, we are in a good position to address the overall question of our investigation, as follows: High levels of ICT capabilities, especially in the areas of Affordability Readiness, Skills Readiness, Political and Regulatory Environment, and Business Usage are associated with the higher level of public value, and, in turn, with a positive socio-economic outcome. We believe that this answer offers a valuable insight to policy and decision-makers, for it does link investments in ICT and the socio-economic outcomes of such investments by a very logical and transparent “Input → Engine → Output” process.

ICT-Enabled Public Value in SSA

217

Over the course of our inquiry we obtained sufficient evidence to support the existence of the hypothesized conceptual link “ICT Capabilities → Public Value → Socio-Economic Outcomes” – our mission seemed to be accomplished. However, because we also discovered the presence of somewhat counterintuitive association rules of the type “High ICT Capabilities → Low Public Value” and “Low ICT Capabilities → High Public Value”, we would like to reconcile this finding with the overall results of the study. The key to an explanation, in our view, can be found in the metadata of WGI. Specifically, WGI relies “exclusively on perceptions-based governance data sources” (Kaufmann et al., 2010c, p. 5) which “include surveys of firms and households, as well as the subjective assessments of a variety of commercial business information providers, non-governmental organizations, and a number of multilateral organizations and other public-sector bodies” (ibid., p. 5). This is consistent with the perspective on public value being defined by the consumers of the value and not the producers of the value (Alford & Hughes, 2008; Cordella & Bonina, 2012). Keeping this in mind, we would like to offer to our readers a set of propositions on which we will base our rationalization for the presence of seemingly counterintuitive association rules. The propositions, presented in the form of a simple list, are as follows: 1. Increase in ICT Capabilities brings about an increase in ICT-related sophistication of the population. 2. A higher level of ICT sophistication of the population brings about a higher level of scrutiny of ICT-enabled products and services (e.g., government’ public services) offered to the population. 3. A lower level of ICT sophistication of the population entails a lower level of scrutiny of ICT-enabled products and services offered to the population. 4. Resultantly, a population with a higher degree of ICT sophistication will apply, ceteris paribus, stricter criteria to appraise ICT-enabled products and services then a population with a lower degree of ICT sophistication. 5. The use of stricter assessment criteria will result, ceteris paribus, in lower assessment scores, and the use of softer assessment criteria will result in higher assessment scores. The presented above five points are easy to follow and to critique – we invite our readers to do so. However, in our view the points above do explain why SSA economies with higher levels of ICT Capabilities appraise Public Value as being lower in quality as compared to their counterparts with lower levels of ICT Capabilities. This explanation is an intuitive one – let us consider, for example, a car mechanic, a person who is very familiar with car’s engine and drivetrain. When purchasing a car, the mechanic’s assessment of a vehicle will be more stringent, and will be performed by using harsher criteria, than the evaluation of a vehicle by a casual user. Simply put a common expression “familiarity breeds contempt”, or as in the case of this study, “familiarity with ICT breeds contempt for ICT-enabled services” may not be off-mark.

218

Quantitative Methodologies using Multi-Methods

CONCLUSION At this point there seem to be a tendency (Cordella & Bonina, 2012) to assess the impact of ICT initiatives in public sector in a fairly narrow, efficiency-based way (Kelly, Mulgan, & Muers, 2002). Unsurprisingly, it was noted that ICT-enabled initiatives in public sector should be evaluated on a broader scale than that of pure efficiency and effectiveness of the target processes (Bannister, 2007; Cordella & Bonina, 2012). In our investigation we expanded the horizon of the impact of ICT capabilities on public services to include socio-economic outcomes, and this we consider to be one of the contributions of our inquiry, for we obtained empirical evidence substantiating the conceptualized chain of links connecting ICT Capabilities, Public Value, and Socio-Economic Outcomes. It has also been shown that the results of ICT-driven interventions could go beyond straight forward improvements in efficiency and effectiveness of the target processes and, instead (or, additionally, for Business Process Reengineering (transformative change) and Total Quality Management (incremental change) are not mutually exclusive), bring about transformative, qualitative changes. And, expectedly, the importance to consider ICT interventions in public service as carrying “potential consequences of the transformation of the relationship between the citizens and the state” has been duly noted (Cordella & Bonina, 2012, p. 515). Indeed, the results of our investigation suggest that ICT can have a transformative impact on the relationship between citizens and the creator of the public value. It seems that the growth in ICT Capabilities tend to bring about the re-assessment of public value generated by public services. Our findings suggest that while citizenry of the economies with low levels of ICT Capabilities may generously appraise the public services as being of high quality, this appraisal will turn more critical as a result of a greater level of sophistication of the populace that was brought about by increases in ICT Capabilities. We consider the framework of this investigation to be another contribution of our effort. Finally, we can claim a contribution to practice in the area of IT4D, for we obtained actionable insights regarding the existence of the condition-dependent relationships between ICT Capabilities, ICT-enabled public value, and socio-economic outcomes.

ACKNOWLEDGMENT Material in this chapter previously appeared in: Representation matters: An exploration of the socio-economic impacts of ICT-enabled public value in the context of sub-Saharan economies. International Journal of Information Management, 49, 69–85 (2019).

REFERENCES Alford, J. and Hughes, O. (2008). Public Value Pragmatism as the Next Phase of Public Management. American Review of Public Administration, 38(2), 130–148. Bannister, F. (2007). The Curse of the Benchmark: An Assessment of the Validity and Value of e-Government Comparisons. International Review of Administrative Sciences, 3(2), 612–631.

ICT-Enabled Public Value in SSA

219

Bannister, F. and Connolly, R. (2014). ICT, Public Values and Transformative Government: A Framework and Programme for Research. Government Information Quarterly, 31(1), 119–128. Bekkers, V. and Homburg, V. (2007). The Myths of E-Government: Looking Beyond the Assumptions of New and Better Government. !e Information Society, 23(5), 373–382. Blaug, R., Horner, L., and Lekhi, R. (2006). Public Value, Politics and Public Management: A Literature Review. Work Foundation, London. Chipman, H. and Tibshirani, R. (2006). Hybrid Hierarchical Clustering with Applications to Microarray Data. Biostatistics Advance Access, 7(2), 286–301. Cordelia, A. (2006). Transaction Costs and Information Systems: Does IT Add Up?. Journal of Information Technology, 21(3), 195–202. Cordelia, A. (2007). E-government: Towards the e-Bureaucratic Form?. Journal of Information Technology, 22(3), 265–274. Cordella, A. and Willcocks, L. (2010). Outsourcing, Bureaucracy and Public Value: Reappraising the Notion of the “Contract State”. Government Information Quarterly, 27(1), 82–88. Cordella, A. and Bonina, C. (2012). A Public Value Perspective for ICT Enabled Public Sector Reforms: A Theoretical Reflection. Government Information Quarterly, 29(4), October 2012, 512–520. Dutta, S., and Jain, A. (2003). The Networked Readiness of Nations. In The Global Information Technology Report 2002–2003, Dutta, S., A. Lanvin, B., & Paua, F. (eds.). Oxford University Press, New York, NY, Oxford. Eppler, M. J. (2007). Information Quality in Electronic Government: Toward the Systematic Management of High-Quality Information in Electronic Government-to-Citizen Relationships. In Governance and Information Technology: From Electronic Government to Information Government, Mayer-Schönberger, V. and Lazer, D. (eds.), The MIT Press, Cambridge, MA. Dunleavy, P., Margetts, H., Tinkler, J., and Bastow, S. (2006). Digital Era Governance: IT Corporations, the State, and e-Government. Oxford University Press, Oxford, UK. Gupta, B., Dasgupta, S., and Gupta, A. (2008). Adoption of ICT in a Government Organization in a Developing Country: An Empirical Study. The Journal of Strategic Information Systems, 17(2), 140–154. Habtamu, F. N. (2008). Roles of Governance in Explaining Economic Growth in Sub-Saharan Africa. Africa Policy Journal, 4, 1–21. Hood, C. and Lodge, M. (2006). The Politics of Public Service Bargains. Oxford University Press, Oxford. Jaeger, P. (2005). Deliberative Democracy and the Conceptual Foundations of Electronic Government. Government Information Quarterly, 22(4), 702–719. Kaufmann, D., Kraay, A., and Mastruzzi, M. (2007). Answering the Critics. World Bank Policy Research Department Working Paper No. 4149. Washington, DC. Kaufmann, D., Kraay, A., and Mastruzzi, M. (2010a). The Worldwide Governance Indicators: a Summary of Methodology, Data and Analytical Issues. World Bank Policy Research Working Paper no. 5431. Available at SSRN: http://ssrn.com/abstract=1682130. Kaufmann, D., Kraay, A., and Mastruzzi, M. (2010b). Response to “What do the Worldwide Governance Indicators Measure?” The European Journal of Development Research, 22(1), 55–58. Kaufmann, D., Kraay, A., and Mastruzzi, M. (2010c). The Worldwide Governance Indicators: Methodology and Analytical Issues. World Bank Policy Research Working Paper 5430. Available on-line at: https://www.brookings.edu/wp-content/uploads/2016/06/09_wgi_ kaufmann.pdf. Kelly, G., Mulgan, G., and Muers, S. (2002). Creating Public Value: An Analytical Framework for Public Service Reform. Strategy Unit of Cabinet Office, London, UK.

220

Quantitative Methodologies using Multi-Methods

Kimaro, E. L., Keong, C. C., and Sea, L. L. (2017). Government Expenditure, Efficiency and Economic Growth: A Panel Analysis of Sub-Saharan African Low Income Countries. African Journal of Economic Review, 5(2), 34–54. Kirkman, S.G., Osorio, A.C., and Sachs, D.J. (2002). The Networked Readiness Index: Measuring the Preparedness of Nations for the Networked World. In The Global Information Technology Report 2001–2002 Readiness for the Networked World, Kirkman, G. (ed.), Oxford University Press, New York, NY, pp. 10–29. Laan, M. and Pollard, K. (2002). A New Algorithm for Hybrid Hierarchical Clustering with Visualization and the Bootstrap. Journal of Statistical Planning and Inference, 117 (2), 275–303. Madon, S., Sahay, S., and Sudan, R. (2007). E-Government Policy and Health Information Systems Implementation in Andhra Pradesh, India: Need for Articulation of Linkages Between the Macro and the Micro. The Information Society, 23(5), 327–344. Moore, M. (1995). Creating Public Value: Strategic Management in Government. Harvard University Press, Cambridge, MA. Samoilenko, S. and Osei-Bryson, K.M. (2008). Increasing the Discriminatory Power of DEA in the Presence of the Sample Heterogeneity with Cluster Analysis and Decision Trees. Expert Systems with Applications, 34(2), 1568–1581. Samoilenko, S. and Osei-Bryson, K.M. (2017). A Methodology for Identifying Sources of Disparities in the Socio-Economic Impacts of ICT Capabilities in Sub-Saharan Economies. In Proceedings of the International Conference on Information Resource Management, Santiago, Chile, May 16–19, 2017. The World Bank (2016). Data Bank: Worldwide Governance Indicators. Available on-line at: http:// databank.worldbank.org/data/reports.aspx?source=worldwide-governance-indicators. Thomas, M. A. (2010). What Do the Worldwide Governance Indicators Measure? European Journal of Development Research, 22, 31–54. WEF_GITR. (2015). World Economic Forum’ Global Information Technology Report. Available on-line at: http://reports.weforum.org/global-information-technology-report-2015/ network-readiness-index/. WEF_NRI (2016). Networked Readiness Dataset. Available on-line at: http://reports.weforum.org/global-information-technology-report-2016/networked-readiness-index/.

25

Contributing Factors to Information Technology Investment Utilization in Transition Economies An Empirical Investigation

INTRODUCTION What are the transition economies (TEs) and where, in the economic map of the world, do they fit? It is not an easy task to categorize the economies of the world according to the single taxonomy; for no single commonly agreed upon taxonomy exists. Perhaps, multiple perspectives shall do a better job of outlining the context of this study. The World Bank classified all economies of the world based on the gross national product (GNP) per capita of a given country in 1999. If the value was less than $755, the country was considered a low-income country. If the value was above $9,266, then the country was labeled as a high-income country, and if the value falls somewhere in between, then it is a middle-income country (World Bank 2002). Middle-income countries can be further sub-divided into two sub-groups of lower-middle-income countries, where GNP per capita is between $756 and $2,995, and upper-middleincome countries, where GNP per capita is between $2,996 and $9,265. The United Nations designates those low-income economies that are characterized by weak human assets (as measured through a composite Human Assets Index) and economic vulnerability as least-developed countries (UNCTAD, 2004) while conceding that no agreed upon criteria exists for categorizing either developed or developing countries (UNSD, 1999). According to the perspective of the World Bank, however, developed countries are the high-income economies that have a high standard of living (often represented in terms of Human Development Index; UNDP, 1998), while developing countries are the low- and middle-income economies that have a low to moderate standard of living. The term transition economies is used to refer to countries in the process of transitioning from a government or a state-controlled centrally planned economy (Ollman, 1997; Myers, 2004) to a market-oriented economy, where the market, rather than government or state, plays the “invisible hand” (Smith, 1776). It does not mean, however, that TEs constitute a homogenous group in terms of the level of economic development. The World Bank, for example, may group some of them with the developed and some with the developing countries, depending on the level of industrialization. 221

222

Quantitative Methodologies using Multi-Methods

Hoskisson et al. (2000) combined two groups of “51 high-growth developing countries in Asia, Latin America, and Africa/Middle East, and 13 transition economies in the former Soviet Union” into the category of emerging economies. They define “emerging economy” as a country that “satisfies two criteria: a rapid pace of economic development, and government policies favoring economic liberalization and the adoption of a free market system”. They warn, “at present, there is no standard list of countries agreed to be emerging economies”, nor that there exists a common agreement on the meaning of the term; consequently, the “term ‘emerging market economy’ may also mean different things to different researchers”. At this point, there is no single accepted taxonomy to classify economies of the world. There are differences between the countries that defy a single framework; nevertheless, there are common traits among them as well. One such trait is that the great majority of the countries in the world invest in information and communication technologies (ICT). Revenue generation serves as a major means by which investments in ICT contribute to macroeconomic growth (UN ICT Task Force Report, 2005; WT/ICT Development Report, 2006). Consequently, improving the effectiveness and efficiency of revenue production is a possible route to increase the macroeconomic impact of investments in ICT. And while these investments have been consistently contributing to economic growth by producing significant and reliable streams of revenues in developed countries, the result of such investments is not as clear-cut in the context of the TEs. In order to make investments in ICT attractive to domestic and international investors, however, TEs must be able to demonstrate their ability to produce revenues from such investments in a reliable and efficient manner. There is little doubt that investments in ICT could and do produce robust returns and contribute to the overall economic growth in the context of developed economies (OECD, 2005; Jorgenson, 2001; Jorgenson & Stiroh, 2000; Oliner & Sichel, 2002; Stiroh, 2002; Colecchia & Schreyer, 2001; Van Ark et al., 2002; Daveri, 2002; Jalava & Pohjola, 2002). In the contexts of developing countries and economies in transition, however, the levels of the returns on investments differ significantly. The TEs present a particularly interesting case for the research because they share economic characteristics with both developed and less developed economic regions (OECD, 2004). For example, domestic markets of TEs “include both substantial populations with significant disposable income and large numbers of people without” (ibid, p. 12). As a result, TEs present a good vantage point from which the relationship between the investments in ICT and economic growth can be investigated. Multiple research studies conducted in this area have identified a group of factors that affect the return on investments in ICT. It has been suggested that the differences in capital stock and infrastructure (Dewan & Kraemer, 2000; Piatkowski, 2002), a level of investments in ICT (Murakami, 1997; Piatkowski, 2002), as well as the amount and quality of the available human capital (OECD, 2004), are some of the variables that impact the level of returns on investments in ICT. We point the attention of the reader to another, rather obvious, but overlooked, determinant of effective production of revenue, namely, the efficiency of utilization of investments. As mentioned above, there is a consensus in the research community

Information Technology Investment Utilization

223

that capital stock and infrastructure, human capital, and level of investments in ICT are factors that affect the economic outcome of investments. However, do these factors also affect the efficiency of utilization of investments? It would appear to be the case in the context of developed countries, but, to our knowledge, there have been no studies undertaken to answer this question in the context of TEs. Furthermore, as economies gradually acquire better infrastructures, invest more in ICT, and improve the level of human capital, do these economies obtain higher levels of returns on investments because of the corresponding gradual increases in efficiency as well? We could not find any reported evidence in the published literature regarding this question either. Let us elaborate on the importance of this research problem a little further. It is reasonable to propose that investments in ICT must be gauged in accordance with the level of efficiency with which these investments can be utilized. The simple supporting reason for this statement is that efficiency is a relative term. It is relative to the external context (i.e., how efficient is country A relative to country B?), as well as to the internal context, and it is the internal context that deals with the level of the available resources. If country A can efficiently utilize X level of the investments in ICT, in the absence of perfect scalability, the efficiency would likely change if the level of investments in ICT drastically changes. However, the issue of the relationship and the interplay between the level of investments in ICT and the efficiency of utilization of investments are simply too complex to be tackled in this study, for there are smaller questions that we must answer first. One such question concerns the factors that affect the level of efficiency of utilization of investments in ICT. In line with related research in this area, this study too is ultimately concerned with the economic outcome of investments in ICT. Unlike the other studies, however, this research inquires specifically into the efficiency of the process by which investments in ICT are utilized, as well as into factors that possibly affect the level of efficiency. More formally, in this research, we look at a subset of investments in ICT, specifically investment in telecoms, and investigate how efficiently TEs utilize these investments to produce revenues, and what factors contribute to the efficiency of investment utilization. For the purposes of our study, we adapt the definition of investment in telecoms provided by the Yearbook of Statistics (2004) (see more about this source in the next section), namely, as an investment that: refers to expenditure associated with acquiring the ownership of telecommunication equipment infrastructure (including supporting land and buildings and intellectual and non-tangible property such as computer software). These include expenditure on initial installations and on additions to existing installations.

In order to address this question, we employ a three-phase methodology utilizing data envelopment analysis (DEA) (Charnes, Cooper, Lewin, & Seiford, 1994), cluster analysis (CA) (Aldenderfer & Blashfield, 1984), and decision trees (DT) (Breiman, Friedman, Olshen, & Stone, 1984). First, we discuss the theoretical framework that supports our inquiry.

224

Quantitative Methodologies using Multi-Methods

THEORETICAL FRAMEWORK Growth Accounting To approach our research problem, we rely on a neoclassical framework of growth accounting. This framework originated from the work of Solow (1957) and since then has been widely used by other researchers (Oliner & Sichel, 2002). The objective of growth accounting is to decompose, using a neoclassical production function, the rate of growth of an economy into the contributions from the different inputs. A neoclassical production function relates output and inputs in the following manner:

Y = f ( A, K , L )

where Y = output (most often in the form of GDP), A = the level of technology/total factor productivity (TFP), K = capital stock, and L = quantity of labor/size of labor force. Which, in the case of this study, becomes: Y = revenues from telecoms; A = TFP; K = investments in telecoms; and L = number of full-time telecom employees. Based on the function provided above, growth accounting uses the Cobb–Douglas production function:

Y = A × K α × Lβ

where α and β are constants determined by technology. In the case of constant returns to scale (e.g., if α + β > 1, then returns are increasing to scale, and if α + β < 1, then returns are decreasing to scale), α + β = 1, thus, β = 1 − α, which gives the following formulation:

Y = A × K α × L1−α

It is important to note that this function does not necessarily represent the true relationships between the inputs and the output; rather, its purpose is simply to serve as a vehicle for exploration and interpretation of the macroeconomic outcomes. Out of three inputs used by growth accounting, only capital K and labor L could be observed in the data, while TFP would serve as a residual (often referred to as Solow’s residual) term capturing that contribution to Y (GDP or revenues from telecoms), which is left unexplained by the inputs of capital and labor. In the case of this study, we are interested, first, in the efficiency of the utilization of the investments in telecoms by full-time telecom employees and, second, in some of the factors that possibly contribute to the level of efficiency. Consequently, for every point in time (for every year in our case) neoclassical production function allows us to relate investments in telecoms, full-time telecom employees, and revenues from telecoms in the following fashion: Revenues from telecoms = f (TFP, investments in telecoms, full-time telecom employees)

Information Technology Investment Utilization

225

As a result, for each TE in the study, for the period of, let us say, ten years we have ten values of the relative efficiency of utilization of investments in telecoms by full-time telecom staff. This approach of determining efficiency at a point in time, rather than determining the change in efficiency over a period of time, has two advantages. First, we do not need to account for depreciation of the telecom infrastructure, assuming the same rate of depreciation for all TEs in the study. Second, because the number of full-time telecom employees is reported annually, it allows us to treat the number of full-time telecom employees as being constant over the period of one year. One of the appeals of using the neoclassical growth accounting framework lies in its simplicity; after all, only two factors, the TFP growth and the rate of increase in inputs, are used to explain the growth rate of the output. As a result, this relationship reflects the fundamental assumptions of the framework, namely, the presence of technological progress and the growth of labor. However, the flip side of this simplicity is the somewhat limited explanatory capability of economic growth. For example, while assuming technological progress the framework neither explains the sources of the progress or the factors that affect the progress nor does it account for any possible interactions between the technological progress and capital growth. In reality, though, capital investments would be affected by the technological progress, for progress in information technologies have fueled capital investments in the economies of the United States and other developed countries. Finally, according to another assumption of the growth accounting framework, namely that the capital is subject to the law of diminishing returns, the convergence of the poor and wealthy economies must take place. The reality, however, reflects that the gap between poor and rich countries of the world is widening. Nevertheless, the use of the growth accounting framework for the purposes of researching contributions of ICT investments to macroeconomic growth of TEs appears to be warranted, for this analytical framework has been widely used to estimate the contribution of ICT to economic growth in the context of developed and developing countries (Oliner & Sichel, 2000; Schreyer, 2000; Daveri, 2000; Jorgenson & Stiroh, 2000; Whelan, 1999; Hernando & Nunez, 2002). Rarely, however, should any type of investment made on the macroeconomic level be perceived and considered in isolation. Rather, it is more beneficial to search and identify a set of complementary factors that allow magnifying the potential benefits of the investment. Often, such complementary factor, or a set of factors, is represented by other types of investments, sometimes in the entirely different areas of the economy. In this chapter, we use theory of complementarity as a theoretical framework that supports our search for the factors that may contribute to the efficiency of utilization of investments in telecoms by full-time telecom employees. A brief overview of the theory of complementarity is offered next.

Theory of Complementarity Initially introduced in economics by Edgeworth (1881), the concept of complementarity refers to the notion that the increase in one factor may result in the increased benefit received from its complementary factors. We apply the theory of

226

Quantitative Methodologies using Multi-Methods

complementarity to our research problem to argue that if the benefits of the investments in telecoms and in full-time telecom labor are to be reaped successfully at the macroeconomic level, then such investments should not be made in isolation from investments in other areas. Thus, if two factors are more effective when taken jointly, rather than separately, we consider such factors complementary. We suggest that the process of identification of complementary factors requires a two-step approach. First, we need to identify a pool of possibly complementary factors. Second, we then proceed further and test those factors for complementarity with the variable of interest. In the case of our study, we are interested in finding a set of factors that may contribute to the efficiency of the utilization of the investments in telecoms by fulltime telecom employees. Consequently, our research can be perceived as aiming to accomplish the first part of the two-step approach. Namely, the aim of our study is to identify a pool of the factors that are possibly complementary to the number of full-time telecom employees. However, even if the complementarity of the investments exists within a given production function, it cannot be identified through the formulation offered by Cobb– Douglas production function. Complementarity of the investments can only be discerned if the formulation allows for the presence of the interaction term between the specified investments. Thus, we turn our attention to the transcendental logarithmic production function, a brief overview of which is offered next. Standard Cobb–Douglas production function, as it was mentioned before, can be represented as:

Y = A × K α × Lβ

By taking the logarithm, the following formulation is obtained:

log Y = log A + α log K + β log L

Extension to the above formulation of the Cobb–Douglas production function, called the transcendental logarithmic (translog) production function, is provided below: log Y = β 0 + β1 × log K + β 2 × log L + β3 × log K 2 + β 4 × log L2 + β5 × log K × log L + e It is easy to see that the Cobb–Douglas function is “nested” in the translog function, and testing whether both functions describe the production process equally well would entail testing the following null hypothesis:

H 0 : β3 = β 4 = β5 = 0

The translog production function is more flexible than the Cobb–Douglas function in the sense that it allows testing for the presence of the interactions between the

Information Technology Investment Utilization

227

variables, where the test for the presence of the interaction would involve testing of the following hypothesis:

H 0 : β5 is not statistically discernible from 0 at the given level of α

Complementarity of investments has been investigated in the context of research and development (R&D) portfolios by Lambertini (2003), Lin and Saggi (2002), Rosenkranz (2003) and in the context of process and product innovation by Athey and Schmutzler (1995). In a more relevant context to this research, Giuri, Torrisi, and Zinovyeva (2005) explored the complementarity between skills, organizational change, and investments in ICT. Bugamelli and Pagano (2004) studied the complementarity between investment in ICT and the related investment in human and organizational capital. Gera and Wulong (2004) examined the complementarity of the investment in ICT and organizational changes and worker skills, and Loukis and Sapounas (2004) inquired into the complementarity between IS investment and the set of IS management factors.

OVERVIEW ON THE DATA In our choice of data analytic tools, we were restricted by our selection of the data which represents a sample of convenience. Use of any parametric method would have required an assumption of data normality, which a sample of convenience may not satisfy. Thus, we decided to use DEA, CA, and DTs, all well-known and dependable non-parametric methods. The data for this study were obtained from the World Development Indicators database (web.worldbank.org/WBSITE/EXTERNAL/DATASTATISTICS), which is the World Bank’s (web.worldbank.org) comprehensive database on development data, and the Yearbook of Statistics (2004) (www.itu.int/ITU-D/ict/publications), which is published yearly by International Telecommunication Union (ITU) (www. itu.int). In our choice of variables, we were greatly restricted by the availability of the data. For example, while the development data of the World Bank’s database covers more than 600 indicators for 208 economies, data on many of the indicators relevant to our research were not available, or were available only for a few countries, or contained too few data points to be useful in data analysis. In our choice of TEs to include in the study, we tried to identify and select a group of countries that started the transition process at approximately the same time. Thus we decided on the following 18 transitional economies: Albania, Armenia, Azerbaijan, Belarus, Bulgaria, Czech Republic, Estonia, Hungary, Kazakhstan, Kyrgyzstan, Latvia, Lithuania, Moldova, Poland, Romania, Slovakia, Slovenia, and Ukraine. In terms of the length of the time series, we were restricted to the period from 1992 to 2004, for which data were provided by the Yearbook of Statistics of ITU. We decided to begin our analysis with the year 1993 because we believe that year provides a common starting point for the transitional economies. Our reasoning here is that it took a year from the dissolution of the Soviet bloc in 1991 for the transition process to begin, and using the year 1992 as a starting point may favor “early starters”.

228

Quantitative Methodologies using Multi-Methods

METHODOLOGY: SEARCHING FOR THE DETERMINANTS OF THE EFFICIENCY OF UTILIZATION OF INVESTMENTS IN TELECOMS In this part of the chapter, we describe, in step-by-step fashion, the methodology used in this study.

Phase 1: Data Envelopment Analysis The cornerstone of our approach is DEA which we utilize to obtain the relative efficiency score for each TE in our sample. Our data set spans a 10-year period, consequently, we perform DEA ten times, ones for each year for the period from 1993 to 2002. As a result, we obtain ten scores of the relative efficiency, one for every year, for each TE in our sample. These scores refer to the relative efficiency of transforming investments in telecoms into revenues from telecoms. From the set of available data, we select a subset representing inputs, and a subset representing outputs. These are the variables that are used in the specification of the DEA model. Data Used to Perform DEA For the DEA part of the methodology, we have identified a model consisting of six input and four output variables. We present the description of the model first and then follow with the justification of the variables that comprise our model. Input variables of the DEA model: 1. GDP per capita (in current US $) 2. Full-time telecommunication staff (% of the total labor force) 3. Annual telecom investment per telecom worker 4. Annual telecom investment (% of GDP in current US $) 5. Annual telecom investment per capita 6. Annual telecom investment per worker Output variables of the DEA model: 1. Total telecom services revenue per telecom worker 2. Total telecom services revenue (% of GDP in current US $) 3. Total telecom services revenue per worker 4. Total telecom services revenue per capita The main goal that we pursue in performing DEA is to find out how efficient the 18 transition economies are in converting investment inputs into the revenue outputs. Therefore, we did not include any other types of inputs or outputs such as those related to infrastructure, capabilities, and utilization. It is should be mentioned that the purpose of our DEA model is not to reflect the path by which the investments are transformed into the revenues over the course of one year, rather, the intent of our model is to depict a “fiscal efficiency” of the TEs regarding their investments in telecoms. Upon the close inspection of the chosen variables, one can see that all of them are expressed as ratios. We intentionally present the levels of investments and revenues

Information Technology Investment Utilization

229

not in absolute dollar terms, but in relative units. The intent in doing so is to lessen the impact of the differences between TEs in terms of their size, population, level of wealth while representing the investments and revenues more broadly (i.e., relative to the whole population, labor force of a country, and the telecom industry). We argue that such relative representation provides a more objective depiction of not only the investments and revenues themselves but also of the economic and demographic environment within which the investments take place and the revenues produced. There are no objective criteria according to which the “best” DEA model can be constructed; instead, the decision about including input and output variables is usually delegated to the purview of the investigator. We would like to, however, provide some justification regarding our choice of the input and output variables in our DEA model. We include the input variable GDP per capita (in current US $) in order to take into consideration the differences between the levels of the economic development of eighteen TEs in the study. Let us recall, that according to the World Bank, economies can be classified as low-income, lower-middle-income, upper-middle-income, and high-income. Consequently, the inclusion of the input variable GDP per capita (in current US $) allows us to account for the possible differences in the level of industrialization of these countries. The reason for the inclusion of the input variable full-time telecommunication staff (% of the total labor force) is intuitive; according to the assumption of the study, investments in telecoms are converted into revenues by full-time telecom employees, who represent one of the essential input components of the revenue-generating process (i.e., without employees, investments cannot be converted into revenues). The inclusion of the rest of the input and output variables of the DEA model is based on the theoretical framework used in our study, neoclassical growth accounting. Again, the reason for representing the variables annual telecom investments and total telecom services revenues relative to GDP, total population, total labor force, and total telecom employees, was to counter the differences between TEs in terms of their size, population, and level of wealth. At the same time, such an approach allows us to obtain some sort of representation of the structure of the economies within which investments convert to revenues.

Phase 2: Cluster Analysis In the second phase of our inquiry, we use CA to determine whether the TEs in our sample are similar in terms of their relevant characteristics, as represented by the input and output variables of the DEA model. Thus, more formally, in this part of the study, we test the null hypothesis that there are no discernable clusters of TEs with respect to their level of investments in and revenues from ICT. This hypothesis can be stated as follows: H0: The sample of 18 TEs is homogenous in terms of the levels of annual telecom investments and total telecom services revenues Consequently, we reject the null hypothesis if CA results in more than one cluster and, given a set of data points representing a transitional economy over a 10-year period of time, every cluster contains a complete set of data points representing a given economy. If the results of CA reveal the presence of multiple subgroups in the sample, then we calculate the averaged over 10 year-period relative efficiency scores (identified in

230

Quantitative Methodologies using Multi-Methods

Phase 1) for each group. We can expect that if heterogeneous subgroups are identified in our sample, then these subgroups have different average relative efficiencies. Data Used to Perform CA To perform CA, we reduce the data set used to conduct DEA by removing the GDP per capita and full-time telecommunication staff variables. While these two variables are important as the inputs of the DEA model, in CA we aim to test homogeneity of the sample in terms of the investments and revenues only, thus, we decided against using these variables in CA. The complete list of the variables used to perform CA is provided below. Variables used to perform CA: 1. Annual telecom investment per telecom worker (Current US $) 2. Annual telecom investment (% of GDP) 3. Annual telecom investment per capita (Current US $) 4. Annual telecom investment per worker (Current US $) 5. Total telecom services revenue per telecom worker (Current US $) 6. Total telecom services revenue (Current US $) 7. Total telecom services revenue per worker (Current US $) 8. Total telecom services revenue per capita (Current US $)

Phase 3: Decision Tree In the third phase of our study, we use DT analysis to identify the most important dimensions that differentiate the heterogeneous subgroups in our sample. The goal of this phase is to identify some of the variables that may be responsible for the differences in average relative efficiencies across the subgroups. In order to do so, we create a new categorical target variable with its domain of values equal to the number of subgroups identified in Phase 2. Once the first, most important split is made, we record the name of the variable and the value at which the split was made. After that, we remove that variable from further analysis and repeat the procedure. Data Used to Perform DT Before conducting the DT analysis, we identified the largest set of data available to us; in our analysis, we were able to use 34 variables, which are listed below. Let us recall, that in this study we are interested in finding a set of factors that may contribute to the efficiency of the utilization of the investments in telecoms by full-time telecom employees. Consequently, the aim of this part of the data analysis is to identify a pool of factors that are possibly complementary to the number of full-time telecom employees. Variable used for DT analysis: 1. Exports of computer, communications and other services (% of commercial service exports) 2. High-technology exports (% of manufactured exports) 3. Imports of computer, communications and other services (% of commercial service imports)

Information Technology Investment Utilization

231

4. Military expenditure (% of GDP) 5. Military personnel (% of the total labor force) 6. Fixed line and mobile phone subscribers (per 1,000 people) 7. International telecom, outgoing traffic (minutes per subscriber) 8. Internet users (per 1,000 people) 9. Mobile phones (per 1,000 people) 10. Telephone mainlines (per 1,000 people) 11. Telephone mainlines per employee 12. Health expenditure per capita (current US $) 13. Health expenditure, private (% of GDP) 14. Health expenditure, public (% of GDP) 15. Health expenditure, total (% of GDP) 16. Immunization, DPT (% of children ages 12–23 months) 17. Immunization, measles (% of children ages 12–23 months) 18. Pupil-teacher ratio, primary 19. School enrollment, secondary (% gross) 20. School enrollment, tertiary (% gross) 21. Research and development expenditure (% of GDP) 22. Researchers in R&D (% of the total labor force) 23. Technicians in R&D (% of the total labor force) 24. Roads, paved (% of total roads) 25. Roads, total network (km) 26. Full-time telecommunication staff (% of the total labor force) 27. Annual telecom investment (% of GDP in current US $) 28. Urban population (% of total) 29. Urban population growth (annual %) 30. Population growth (annual %) 31. Foreign direct investment, net inflows (% of GDP) 32. GDP growth (annual %) 33. GDP per capita (constant 2000 US $) 34. GDP per capita growth (annual %)

RESULTS Results: DEA In this section, we describe the results of the DEA. To perform DEA, we used the software application “OnFront”, version 2.02, produced by Lund Corporation (www.emq.com). In using “OnFront” to obtain the efficiency scores, we have chosen to use Farrel Input-Saving Measure of Efficiency as a direct efficiency measure for the three types of models: CRS (constant return to scale), VRS (variable return to scale), and NIRS (non-increasing return to scale). In Table 25.1, we provide the averaged over 10 years scores of the relative efficiency of the 18 TEs. The complete results of DEA, with the scores for each year, are provided in Appendix A. The results of DEA show that a number of transitional economies in some years obtained a rating of hundred-percent relative efficiency. This, sometimes overly

232

Quantitative Methodologies using Multi-Methods

TABLE 25.1 Averaged Scores of Relative Efficiency, per Country, per DEA Model Country Albania Armenia Azerbaijan Belarus Bulgaria Czech Republic Estonia Hungary Kazakhstan Kyrgyzstan Latvia Lithuania Moldova Poland Romania Slovakia Slovenia Ukraine

CRS 0.98 0.89 0.73 0.51 0.92 0.85 0.96 1.00 0.64 0.93 0.76 0.66 0.92 0.93 0.58 0.81 0.95 0.91

VRS 1.00 0.91 0.97 0.68 0.94 0.90 0.97 1.00 0.77 1.00 0.89 0.83 1.00 0.96 0.70 0.84 0.98 0.93

NIRS 0.98 0.89 0.73 0.51 0.94 0.86 0.96 1.00 0.67 0.93 0.76 0.67 0.92 0.93 0.58 0.81 0.95 0.93

generous assignment of efficiency scores, is a common characteristic of the most DEA models (Lins et al., 2003). Another common characteristic of DEA models is that they tend to evaluate as efficient those DMUs that have the smallest input values, or, the DMUs with the largest outputs (Ali, 1994). Consequently, based only on the results of the DEA analysis we cannot determine the true nature of the relative efficiency of the TEs in our sample. It is possible that the relatively efficient TEs obtained their status because they are indeed more efficient in the utilization of the inputs than the relatively inefficient ones. It is also possible, however, that the status of being relatively efficient was awarded to the TEs with the lowest levels of the investments in telecoms.

Results: Cluster Analysis Let us recall that we conducted CA with the purpose of identifying the presence of possible differences between the TEs in terms of the levels of investments and the revenues. The variables subjected to CA are not measured on the same scale, so, prior to CA the data had to be standardized. We used SAS Enterprise Miner (EM) to perform CA. We started our inquiry by choosing “Automatic” setting, which did not require any input from the investigator regarding the desired number of clusters. The “Automatic” setting of EM uses Standard Least Squares clustering criterion (which minimizes the sum of squared distances of data points from the cluster

Information Technology Investment Utilization

233

TABLE 25.2 Membership of the Two-Cluster Solution Majority Albania (1993–2002) Armenia (1993–2002) Azerbaijan (1993–2002) Belarus (1993–2002) Bulgaria (1993–2001) Estonia (1993) Kazakhstan (1993–2002) Kyrgyzstan (1993–2002) Latvia (1993, 1996) Lithuania (1993–1998) Moldova (1993–2002) Romania (1993–2002) Slovakia (1993, 1994, 1999) Ukraine (1993–2002)

Leaders Bulgaria (2002) Czech Rep (1993–2002) Estonia (1994–2002) Hungary (1993–2002) Latvia (1994, 1995, 1997–2002) Lithuania (1999–2002) Poland (1993–2002) Slovenia (1993–2002) Slovakia(1995–1998, 2000–2002)

means), Ward’s Minimum Variance1 as a clustering method, and limits the minimum number of clusters to 2 and the maximum to 40. By beginning with this setting, which resulted in a five-cluster solution, we were able to determine the starting point in our analysis. By requesting fewer and fewer number of clusters, we then gradually derived four-, three-, and two-cluster solutions. By using CA, we were able to come up with a solution that partitions our data set into two clusters. The membership of each cluster is provided in Table 25.2. One of the clusters contains the data points completely representing Poland, Czech Republic, Hungary, and Slovenia over the 10-year period, while the second cluster contains the data points completely representing Albania, Armenia, Azerbaijan, Belarus, Kazakhstan, Kyrgyzstan, Moldova, Romania, and Ukraine. Thus, these results suggest that we are able to reject the null hypothesis regarding homogeneity of 18 TEs in terms of investments and revenues from telecoms. Once the results of CA were obtained, we separated our data set into the two subgroups, and calculated the scores of the averaged relative efficiency for each cluster. According to these calculations, one of the clusters, members of which include Czech Republic, Hungary, Poland, and Slovenia, has higher averaged relative efficiency scores than the cluster containing Albania, Armenia, Azerbaijan, Belarus, Kazakhstan, Kyrgyzstan, Moldova, Romania, and Ukraine. Subsequently, we call the first group the “Leaders” and the second group the “Majority”. These findings are consistent with the results reported by Piatkowski (2003), who studied eight transition economies of Europe (Bulgaria, Czech Republic, Hungary, Poland, Romania, Russia, Slovakia, and Slovenia) and concluded that in the period 1

SAS System Documentation offers following description: “Ward’s method tends to join clusters with a small number of observations, and it is strongly biased toward producing clusters with roughly the same number of observations. It is also very sensitive to outliers (Milligan, 1980)”.

234

Quantitative Methodologies using Multi-Methods

TABLE 25.3 Comparison of the Clusters Based on DEA Criterion for Comparison Average efficiency score, CRS Average efficiency score, VRS Average efficiency score, NIRS

“Leaders” Cluster 0.89 0.95 0.89

“Majority” Cluster 0.79 0.88 0.80

Difference 0.10 0.07 0.09

Difference (%) 12.54  7.48 11.63

between 1995 and 2000 ICT capital has most potently contributed to output growth in the Czech Republic, Hungary, Poland, and Slovenia. Let us elaborate some on the importance of the results of CA. Hoskisson et al. (2000) state that even within the same geographic region, emerging market economies are not homogenous, and the differences between the countries make comparisons in small samples problematic. The results of CA, first, confirm that our set of 18 TEs is, indeed, not homogenous. Second, the results demonstrate that the heterogeneity of the sample established in this study is very specific to the telecommunication industry. Meaning, it may be possible to produce entirely different groupings in the case of a comparison of 18 TEs in terms of the investments in and revenues from, let us say, international tourism, or research and development. It may even be that the 18 TEs are homogenous in some regard, such as, let us suppose, percentage of paved roads. However, we aimed to test the null hypothesis regarding the homogeneity of 18 TEs in terms of the investments in and revenues from the investments in telecoms, and we rejected it based on the results of CA. For the convenience of the reader, we provide summarized results in Table 25.3.

Results: Decision Tree In this part of our analysis, we use DT to identify the characteristics of those TEs, which are the most efficient in utilizing their investments in telecoms. First, we identified the largest (in terms of the number of variables) set of the data that Yearbook of Statistics and WDI Database could yield (complete list of the variables is provided in Section 3.3.1). Second, we created a binary dummy variable, which was set as a “target” of the DT analysis. We assigned the values of “1” of the target variable to the countries comprising the “Leaders” cluster, and we assigned the values of “0” to the members of the “Majority” cluster. The third step consisted of the iterative generation of multiple DT models, where every iteration involved the following three steps. The first step consisted of generating a DT model. The second step involved identifying the variable that was used for the top split, as well as recording the split value for that variable. During the third step, the top split variable of a current iteration was taken out from the data set (this was obtained by setting its status to “don’t use” in the Decision Tree node of EM). Then the process was repeated. In our evaluation of the resulting models, we were looking for those variables, splits along which resulted in the cleanest possible separation of the data set according to the value of the target variable. Consequently, after analysis of the generated DT

Information Technology Investment Utilization

235

models, we ended up with 15 variables that vividly differentiate the “Leaders” from the “Majority”. Moreover, we have calculated the average value of the split variable, as well as the (approximate) percentile within which the value of the split falls. The results of the DT analysis suggest that on average “Leaders” have higher: • • • • • • • • •

GDP per capita level of annual telecom investment level of international telecom traffic number of mobile phones number of telephone mainlines percentage of the internet users among the population number of teachers per pupil in the system of primary education percentage of the total labor force employed as R&D technicians level of spending on health care

At the same time, “Leaders” have lower than the “Majority”: • level of military expenditure • percentage of the labor force serving in the military. All the compiled information allows us to conclude that the “Leaders” appear to be wealthier, in general, than the “Majority”, having better infrastructure and smaller armies. In addition to providing some new insights, our findings are congruent with the results of the previous research conducted in the context of TEs. Cornia and Popov (2001) state that “socialist economies differed considerably among each other in terms of the military sector” and suggest that the resulting “structural distortions” affect the ability of TEs to sustain output during the transition. Results of our inquiry confirm that higher levels of militarization are associated with lower levels of macroeconomic output. Carment (2001) points out that the excessive military expenditures are associated with the reduction of investments in social sectors. Our findings demonstrate that higher levels of military expenditure of the “Majority” are associated with their lower levels of spending on health care. Campos and Coricelli (2002) state that a “quality of infrastructure fundamental for the functioning of a market economy, proxied by telephone lines” and human capital play an important role in the macroeconomic growth of TEs. Again, our investigation found evidence that TEs with better infrastructure and higher quality of human capital tend to have higher levels of macroeconomic output. Based on the results of the DT analysis, we summarize a list of the contributing factors affecting the efficiency of utilization of investments in telecoms, as well as a per factor distribution of the “Majority” and the “Leaders”, in Table 25.4. We also provide the averages for each group and the levels of split in terms of the absolute values for each variable in Table 25.5. The information that was obtained from the analysis of each of the DT models presented as a graph in Figure 25.1. The list of contributing factors affecting the efficiency of utilization of investments in telecoms yielded by DT analysis can be reduced to a smaller number of general factors. One of the possible groupings is presented in Table 25.6.

236

Quantitative Methodologies using Multi-Methods

TABLE 25.4 List of the Contributing Factors and Corresponding Distribution per Group of TEs Distribution Variable/Contributing Factor GDP per capita(constant 2,000 US $) Annual Telecom Investment (% of GDP) Fixed line and mobile phone subscribers International telecom, outgoing traffic, minutes per subscriber(per 1,000 people) Telephone mainlines(per 1,000 people) Telephone mainlines per employee Health Expenditure, Public (% of GDP) Health Expenditure per Capita(Current US $) Military personnel(% of labor force) Military expenditure (% of GDP) Pupil-teacher ratio(primary) Internet users (per 1,000 people) Mobile phones(per 1,000 people) Technicians in R&D(% of labor force) R&D expenditure(% of GDP)

Majority 97% are in bottom 60% 62% are in bottom 43% 94% are in bottom 75% 88% are in bottom 55%

Leaders 90% are in top 40% 98% are in top 57% 63% are in top 25% 90% are in top 45%

75% are in bottom 43% 65% are in bottom 37% 100% are in bottom 75% 88% are in bottom 50% 72.5% are in top 40% 63% are in top 36% 100% are in bottom 70% 75% are in bottom 45% 60% are in bottom 34% 50% are in bottom 27% 67% are in bottom 45%

100% are in top 57% 100% are in top 63% 67% are in top 25% 100% are in top 50% 100% are in bottom 60% 100% are in bottom 64% 63% are in top 30% 93% are in top 55% 100% are in top 66% 100% are in top 73% 83% are in top 55%

TABLE 25.5 List of the Contributing Factors, Averages in Absolute Values, and Levels of Split Averages (In Absolute Values) Variable/Contributing Factor GDP per capita(constant 2,000 US $) Annual Telecom Investment (% of GDP) Fixed line and mobile phone subscribers (per 1,000 people) International telecom, outgoing traffic, minutes per subscriber Telephone mainlines(per 1,000 people) Telephone mainlines per employee Health Expenditure, Public (% of GDP) Health Expenditure per Capita(Current US $) Military personnel(% of labor force) Military expenditure (% of GDP) Pupil-teacher ratio(primary) Internet users (per 1,000 people) Mobile phones(per 1,000 people) Technicians in R&D(% of labor force) R&D expenditure(% of GDP)

Majority 1034.49    0.60  158.78   95.13  160.46   71.06    3.15   77.05    1.12    2.27   18.57  114.55  329.70    0.06    0.52

Leaders 4636.11    1.11  485.24   98.86  296.63  154.89    5.05  383.27    1.95    1.95   14.22   27.72   73.37    0.11    0.87

Level of Split 2720.51    1.11  351.27  117.08  249.96  122.90    5.00  151.00    1.64    2.09   15.71   22.86   81.03    0.048    0.57

Information Technology Investment Utilization

237

FIGURE 25.1  “Majority” (dark grey) vs. “Leaders” (light grey): comparison in terms of 15 criteria. Legend: “GDP per capita” – 90% of the “Leaders” are in the top 40% while 97% of the “Majority” are in the bottom 60%.

TABLE 25.6 General Factors Contributing to Efficiency of Utilization of ICT Investment in TEs General Factors Level of economic development Level of investments in telecoms Level of accumulated ICT capital Level of utilization of accumulated ICT capital Level of socio-technical development

Level of militarization

Contributing Factor GDP per capita (constant 2,000 US $) Annual telecom investment (% of GDP) Telephone mainlines (per 1,000 people) telephone mainlines per employee mobile phones (per 1000 people) Fixed line and mobile phone subscribers international telecom, outgoing traffic, minutes per subscriber internet users (per 1,000 people) Health expenditure, public (% of GDP) health expenditure per capita (Current US $) pupil-teacher ratio (primary) technicians in R&D (% of labor force) R&D expenditure (% of GDP) Military personnel (% of labor force) military expenditure (% of GDP)

238

Quantitative Methodologies using Multi-Methods

CONTRIBUTION OF THE STUDY This study makes several contributions to the existing body of knowledge. First, our research was driven by accepted and well-established theoretical frameworks, which have previously been successfully applied in the context of developed countries. It is true that in some areas of inquiry, such as research on strategies in emerging market economies, “theories promulgated for developed market economies may not be appropriate for emerging economies” (Hoskisson et al., 2000). However, we demonstrated that the same theoretical frameworks that drive inquiries in the context of developed countries can drive research on the efficiency of the utilization of resources, effectiveness of the production of revenue, and complementarity of the investments, in the context of TEs. Second, our research suggests a theory-driven methodology that can be used for the purpose of identification of the pool of possibly complementary factors. We also outlined how our research can be extended (i.e., via translog function) to include the actual test for interaction between the factors in the study. Third, we corroborate the findings of earlier studies regarding the factors affecting the economic outcomes of investments in ICT in the context of TEs. Namely, we demonstrate that the efficiency of utilization of investments in telecoms, one of the determinants of the economic outcomes of investments, is affected by the factors reflecting the levels of investments in telecoms, level of accumulated telecom capital, and level and quality of the available human capital. Fourth, in addition to corroborating the results of the previous studies, our research obtained some new empirical findings in the form of a set of factors affecting the level of the economic outcome of investments in telecoms. Namely, we were able to demonstrate that the overall level of economic development, the level of utilization of the accumulated telecom capital, and the level of militarization of the economy are among the factors that affect the efficiency of the utilization of investments in telecoms. Fifth, from the theoretical standpoint, this research provides a contribution to the existing body of knowledge by suggesting the additional set of variables that should be included in the model describing the relationship between investments in ICT and the economic outcomes of such investments. This study also sheds some new light regarding the type of complementary investments that must take place in parallel with investments in ICT. The results of the study could be of benefit to the community of practitioners as well. Any policy-maker or investor shall be well served by taking into consideration the multiple factors affecting the economic outcomes of investments in telecoms, acknowledging, therefore, that no significant increase in the level of investments in telecoms shall directly result a similar increase in revenues from such investments. Moreover, the results of the study also suggest that an increase in efficiency of the utilization of the resources should go hand-in-hand with the increase in investments and vice versa. Meaning, any significant increase in investments in telecoms should be accompanied by the increase in the efficiency of utilization of these investments, while it is unlikely that any significant increase in efficiency of utilization of resources can take place without any additional investments.

Information Technology Investment Utilization

239

This bears an important implication on strategy of investments in telecoms in the context of TEs. The highly industrialized and efficient (in terms of the utilization of resources) TEs can handle, effectively and efficiently, one-time large investments. However, the investments in the context of the less industrialized TEs, characterized by inefficient utilization of resources, should be made gradually, in step-by-step fashion. Consequently, the results of each investment step must be evaluated against the criteria of increasing efficiency of utilization of investments before any additional investments are made.

SUMMARY AND CONCLUSION In this study, we searched for some of the factors affecting the efficiency of utilization of investments in telecoms in the context of 18 TEs. The use of DEA allowed us to determine the relative efficiency of the utilization of investments by each TE in the sample. By using CA, we were able to demonstrate the presence of the two heterogeneous subgroups within our sample of 18 TEs. By incorporating the results of CA with the results of DEA, we determined the presence of a significant difference between the two groups of TEs in terms of the relative efficiency of the utilization of investments in telecoms. We named the subgroup with the higher averaged relative efficiency the “Leaders” and the group with the lower averaged relative efficiency the “Majority”. By using DT, we were able to demonstrate that the “Leaders” differs from the “Majority” in terms of 15 factors that reflect the differences in the level of economic development, level of investments in telecoms, level of accumulated telecom capital, as well as the level of utilization of telecom capital, level of socio-technical development, and level of militarization of TE. The results of our study strongly suggest that these factors affect the level of the efficiency of the utilization of investments in telecoms in the context of TEs. Based on the results of our study, we hypothesize that the “majority” has a lower level of relative efficiency of utilization of investments in telecoms because of the three following reasons: first, the “majority”, in comparison with the “leaders”, simply does not invest enough in telecoms to be concerned with the issue of efficiency. Second, even if the “majority” does invest at a relatively sufficient level, a large part of the investments is directed not toward obtaining revenue, but toward building the required supporting infrastructure. Because only a part of the overall investments is involved in revenue production, it is reflected as an overall low-efficiency score. Third, as the level of the telecom infrastructure increases, it brings about an increased complexity associated with the utilization of the infrastructure. Consequently, the process by which investments are transformed into revenues becomes more complicated, and the “majority” lacks the necessary socio-technical “know-how” to manage that increased complexity. Testing of these hypotheses represents the outline of some of the possible directions for future research in this area. Intuitively, it would make sense to expect a gradual increase in investments as being accompanied by an associated gradual increase in the level of learning regarding the utilization of the investments, in the form of the socio-technical “know-how”. However, at this point it is not clear whether the increase in efficiency brings about the increase in investments, or whether the

240

Quantitative Methodologies using Multi-Methods

additional investments allow for obtaining the higher efficiency of the utilization of the investments. Despite the contributions that this study makes, our research is not without its limitations. First, the use of the different clustering criteria may produce a clustering solution that is different from the one obtained in this research. Second, during DEA we used a model where all the input variables and the output variables were weighted equally. However, some of the variables may be more important than other variables and, therefore, such variables perhaps should be weighted more heavily than other variables. Moreover, while we provided a rationalization for the DEA model used in our study, ultimately, there are no objective criteria according to which such model can be constructed. Hence, a different DEA model could result in different scores of relative efficiency. Another limitation of this study is associated with the quantity of the data. Clearly, our research may have offered richer insights if more variables were available from the data sources that we used. This limitation, however, is not unique to our study but is characteristic to the research in this area in general (Hoskisson et al. 2000). In addition, it may be argued that the time series data covering a 10-year period could be insufficient to inquire into the nature of the events taking place on a macro-economic level. Nevertheless, in the area of research where it “would appear to be a major need for longitudinal studies” (Hoskisson et al., 2000), we feel that the contributions provided by our study outweigh its limitations.

ACKNOWLEDGMENT Material in this chapter previously appeared in: Contributing Factors to Information Technology Investment Utilization in Transition Economies: An Empirical Investigation, Information Technology for Development, 14(1), 52–75.

REFERENCES Aldenderfer, M.S., & Blashfield, R.K. (1984). Cluster Analysis. Sage, Newbury Park, CA. Ali, A.I. (1994). Computational aspects of DEA. In: Charnes, A., Cooper, W.W., Lewin, A. & Seiford, L.M. (eds.). Data Envelopment Analysis: Theory, Methodology and Applications. Kluwer Academic Publishers, Boston, MA, pp. 63–88. Athey, S., & Schmutzler, A. (1995). Product and process flexibility in an innovative environment. RAND Journal of Economics, 26(4), 557–574. Breiman, L., Friedman, J.H., Olshen, R.A., & Stone, C.J. (1984). Classification and Regression Trees. Chapman & Hall, New York, NY. Bugamelli, M., & Pagano, P. (2004). Barriers to iin ICT. Applied Economics, Taylor and Francis Journals, 36(20), 2275–2286. Campos, N., & Coricelli, F. (2002). Growth in transition: What we know, what we don’t and what we should. Journal of Economic Literature, American Economic Association, 40(3), 793–836. Charnes, A., Cooper, W. W., Lewin, A. Y., & Seiford, L. M. (Eds.). (1994). Data Envelopment Analysis: Theory, Methodology, and Applications. Kluwer, Boston, MA. Colecchia, A., & Schreyer, P. (2001). The Impact of Information Communications Technology on Output Growth. STI Working Paper 2001/7. OECD, Paris.

Information Technology Investment Utilization

241

Cooper, W.W., Seiford, L.M., & Zhu, Joe. (2004). Data envelopment analysis: history, models and interpretations. In: W.W. Cooper, L.M. Seiford and J. Zhu. (Eds.). Handbook on Data Envelopment Analysis. Kluwer Academic Publishers, Boston, MA, pp. 1–39. Cornia, G.A., & Popov, V. (2001). Structural and institutional factors in the transition to the market economy: An overview. In: Cornia, G.A. & Popov, V. (Eds.). Transition and Institutions: The Experience of Late Reformers. Oxford University Press, New York, NY, pp. 3–29. Daveri, F. (2000). Is Growth an Information Technology Story in Europe too? Working Paper. University of Parma, Parma, Italy. Daveri, F. (2002). The new economy in Europe, 1992–2001. Oxford Review of Economic Policy, 18(3), 345–362. Dewan, S., & Kraemer, K. (1998). International dimensions of the productivity paradox. Communications of the ACM, 41(8), 56–62. Dewan, S., & Kraemer, K. (2000). Information technology and productivity: evidence from country level data. Management Science (special issue on the Information Industries), 46(4), 548–562. Edgeworth, F.Y. (1881). Mathematical Psychics. Kegan Paul, London. Gera, S., & Wulong, Gu. (2004). The effect of organizational innovation and information and communications technology on firm performance. International Productivity Monitor, Centre for the Study of Living Standards, 9, 37–51. Giuri, P., Torrisi S., & Zinovyeva, N. (2005). ICT, Skills and Organizational Change: Evidence from a Panel of Italian Manufacturing Firms. LEM Papers Series 2005/11. Laboratory of Economics and Management (LEM), Sant’Anna School of Advanced Studies, Pisa, Italy. Hernando, I., & Nunez, S. (2002). The Contribution of ICT to Economic Activity: A Growth Accounting Exercise with Spanish Firm-Level Data. Banco de España – Servicio de Estudios Documento de Trabajo No. 0203. Available at: http://www.bde.es/informes/ be/docs/dt0203e.pdf. Hoskisson, R., Eden, L., Lau, C., & Wright, M. (2000). Strategy in emerging economies. Academy of Management Journal, 43(3), 249–267. Jalava, J., & Pohjola, M. (2002). Economic growth in the new economy: Evidence from advanced economies. Information Economics and Policy, 14(2), 189–210. Jorgenson, D.W. (2001). Information technology and the US economy. American Economic Review, 91(March), 1–32. Jorgenson, D.W., & Stiroh, K.J. (2000). Raising the speed limit: US economic growth in the information age. Brookings Papers on Economic Activity, 2(1), 125–211. Lambertini, L. (2003). The monopolist’s optimal R&D portfolio. Oxford Economic Papers, 55, 561–78. Lin, P., & Saggi, K. (2002). Product differentiation, process R&D, and the nature of market competition. European Economic Review, 46, 201–211. Lins, M.P.E., Gomes, E.G., de Mello, J.C.C.B.S., & de Mello, A.J.R.S. (2003). Olympic ranking based on a zero sum gains DEA model. European Journal of Operational Research, 148(2), 312–322. Loukis, E., & Sapounas, I. (2004). The Impact of Information Systems Investment and Management on Business Performance in Greece. ECIS 2005 Conference paper. Murakami, T. (1997). The Impact of ICT on Economic Growth and the Productivity Paradox. Available on-line at: http://www.tcf.or.jp/data/19971011_Takeshi_Murakami_2.pdf. Myers, D. (2004). Construction Economics. Spon Press, UK. OECD (2004). DAC Network on Poverty Reduction: ICTs and Economic Growth in Developing Countries. OECD, Paris. Available on-line at: http://www.oecd.org/dataoecd/15/54/34663175.pdf.

242

Quantitative Methodologies using Multi-Methods

OECD (2005). Good practice paper on ICTs for economic growth and poverty reduction. The DAC Journal, 6(3). Oliner, S.D., & Sichel, D.E. (2000). The resurgence of growth in the late 1990s: Is information technology story? Journal of Economic Perspectives, 14(4), 3–22. Oliner, S.D., & Sichel, D.E. (2002). Information technology and productivity: Where are we now and where are we going? Economic Review, 3(3), 15–41. Ollman, B. (1997). Market Socialism: The Debate Among Socialists. Routledge, UK. Piatkowski, M. (2002). The ‘New Economy’ and Economic Growth in Transition Economies. WIDER Discussion Paper No. 2002/63. WIDER, Helsinki. Piatkowski, M. (2003). Does ICT Investment Matter for Output Growth and Labor Productivity in Transition Economies? TIGER Working Paper Series, No. 47. Available on-line at: www.tiger.edu.pl. Rosenkranz, S. (2003). Simultaneous choice of process and product innovation when consumers have a preference for product variety. Journal of Economic Behavior & Organization, 50(2), 183–201. Schreyer, P. (2000). The Contribution of Information and Communication Technology to Output Growth: a Study of the G7 Countries. STI Working Papers 2000/2. OECD, Paris. Solow, R. (1957). Technical change and the aggregate production function. Review of Economics and Statistics, 39(3), 312–20. Smith, A. (1776). The Wealth of Nations. Available on-line at: http://www.marxists.org/ reference/archive/smith-adam/works/wealth-of-nations/. Stiroh, K. (2002). Information technology and the U.S. productivity revival: What do the industry data say? American Economic Review, 92(5), 1559–1576. United Nations Conference and Development [UNCTAD]. (2004). The Least Developed Countries Report. United Nations Publications. United Nations Development Program [UNDP]. (1998). Human Development Report. Oxford University Press, New York, NY. UN ICT Task Force Report. (2005). Innovation and Investment: Information and Communication Technologies and the Millennium Development Goals. Report Prepared for the United Nations ICT Task Force in Support of the Science, Technologies and Innovation Task Force of the United Nations Millennium Project. Available on-line at: www.unicttaskforce.org/. United Nations Statistics Division [UNSD]. (1999). Standard Country or Area Codes for Statistical Use. Statistical Papers, Series M, No.49 Rev. 4. United Nations, New York, NY. Van Ark, B., Melka, J., Mulder, N., Timmer, M., & Ypma, G. (2002). ICT Investments and Growth Accounts for the European Union, 1980–2000. Research Memorandum GD-56, Groningen Growth and Development Centre, Groningen. Available on-line at: www.eco.rug.nl/ggdc/homeggdc.html. World Bank. (2002). Global Economic Prospects and the Developing Countries. World Bank. Available on-line at: http://web.worldbank.org/WBSITE/EXTERNAL/EXTDEC/ EXTDECPROSPECTS/GEPEXT/EXTGEP2002. WT/ICT Development Report. (2006). Measuring ICT for Social and Economic Development. International Telecommunication Union’s World Telecommunication/ICT Development Report, 8th edition. Available on-line at: http://www.itu.int/ITU-D/ict/publications/ wtdr_06/index.html. Yearbook of Statistics. (2004). Telecommunication Services Chronological Time Series 1993–2002. ITU Telecommunication Development Bureau (BDT), International Telecommunication Union. Available on-line at: www.itu.int/ITU-D/ict/publications.

243

Information Technology Investment Utilization

APPENDIX A TABLE A.1 Farrel Input-Saving Measure of Efficiency, CRS Country Albania Armenia Azerbaijan Belarus Bulgaria Czech Republic Estonia Hungary Kazakhstan Kyrgyzstan Latvia Lithuania Moldova Poland Romania Slovakia Slovenia Ukraine

1993 1 0.79 0.49 0.56 1 1 0.81 1 0.80 1 0.41 0.82 0.22 0.93 0.55 0.82 1 1

1994 1 1 0.60 0.31 1 1 0.79 1 0.40 1 0.99 0.82 1 1 0.52 0.80 1 0.70

1995 1 1 1 0.68 1 0.78 1 1 0.60 1 0.94 0.74 1 0.90 0.62 0.79 1 0.70

1996 1 1 1 0.64 0.88 0.60 1 1 0.67 0.72 0.94 0.85 1 0.76 0.59 0.87 1 1

1997 1 1 1 0.56 0.86 0.81 1 1 0.75 0.56 1 0.45 1 0.85 0.48 0.55 1 0.92

1998 1 0.89 0.88 0.47 0.83 0.97 1 1 0.65 1 0.85 0.46 1 1 0.46 0.51 1 1

1999 1 0.81 0.51 0.57 0.92 0.98 1 1 0.51 1 0.69 0.50 1 0.84 0.55 0.71 1 1

2000 0.80 0.73 0.84 0.44 1 0.81 1 1 0.54 1 0.67 0.46 1 1 0.52 1 0.71 1

2001 1 0.86 0.49 0.39 0.72 0.68 1 1 0.51 1 0.55 0.55 1 1 0.63 1 0.81 0.85

2002 1 0.80 0.48 0.52 1 0.90 1 1 1 1 0.52 0.96 1 1 0.90 1 1 0.88

1998 1 0.89 0.95 0.60 0.83 0.97 1 1 0.65 1 0.93 0.63 1 1 0.65 0.66 1 1

1999 1 0.83 1 0.64 1 1 1 1 0.70 1 0.90 0.81 1 0.96 0.73 0.74 1 1

2000 0.95 0.75 1 0.67 1 0.95 1 1 0.78 1 0.94 0.91 1 1 0.75 1 0.96 1

2001 1 0.87 1 0.66 0.72 0.74 1 1 0.75 1 0.83 0.86 1 1 0.72 1 0.87 0.86

2002 1 0.80 1 0.77 1 0.94 1 1 1 1 0.76 1 1 1 0.91 1 1 1

TABLE A.2 Farrel Input-Saving Measure of Efficiency, VRS Country Albania Armenia Azerbaijan Belarus Bulgaria Czech Republic Estonia Hungary Kazakhstan Kyrgyzstan Latvia Lithuania Moldova Poland Romania Slovakia Slovenia Ukraine

1993 1 1 0.82 0.75 1 1 0.86 1 1 1 0.57 0.87 1 0.93 0.61 0.84 1 1

1994 1 1 0.92 0.59 1 1 0.83 1 0.48 1 1 0.93 1 1 0.63 0.82 1 0.70

1995 1 1 1 0.77 1 0.80 1 1 0.62 1 1 0.82 1 0.91 0.67 0.80 1 0.75

1996 1 1 1 0.72 0.93 0.75 1 1 0.70 1 0.95 0.87 1 0.85 0.66 0.87 1 1

1997 1 1 1 0.64 0.95 0.87 1 1 1 1 1 0.60 1 0.93 0.63 0.68 1 1

244

Quantitative Methodologies using Multi-Methods

TABLE A.3 Farrel Input-Saving Measure of Efficiency, NIRS Country Albania Armenia Azerbaijan Belarus Bulgaria Czech Republic Estonia Hungary Kazakhstan Kyrgyzstan Latvia Lithuania Moldova Poland Romania Slovakia Slovenia Ukraine

1993 1 0.79 0.49 0.56 1 1 0.81 1 0.80 1 0.41 0.82 0.22 0.93 0.55 0.82 1 1

1994 1 1 0.60 0.31 1 1 0.79 1 0.40 1 1 0.93 1 1 0.52 0.80 1 0.70

1995 1 1 1 0.68 1 0.78 1 1 0.60 1 1 0.74 1 0.90 0.62 0.79 1 0.70

1996 1 1 1 0.64 0.93 0.60 1 1 0.67 0.72 0.94 0.87 1 0.76 0.59 0.87 1 1

1997 1 1 1 0.56 0.95 0.81 1 1 1 0.56 1 0.45 1 0.85 0.48 0.55 1 1

1998 1 0.89 0.88 0.47 0.83 0.97 1 1 0.65 1 0.85 0.46 1 1 0.46 0.51 1 1

1999 1 0.81 0.51 0.57 1 1 1 1 0.51 1 0.69 0.50 1 0.84 0.55 0.74 1 1

2000 0.80 0.73 0.84 0.44 1 0.81 1 1 0.54 1 0.67 0.46 1 1 0.52 1 0.71 1

2001 1 0.86 0.49 0.39 0.72 0.68 1 1 0.51 1 0.55 0.55 1 1 0.63 1 0.81 0.85

2002 1 0.80 0.48 0.52 1 0.90 1 1 1 1 0.52 0.96 1 1 0.90 1 1 1

26

Increasing the Discriminatory Power of DEA in the Presence of the Sample Heterogeneity with Cluster Analysis and Decision Trees

INTRODUCTION Data envelopment analysis (DEA) is a non-parametric data analytic technique that is extensively used by various research communities (e.g., Hong et al., 1999; Sohn and Moon, 2004; Seol et al., 2007) since its introduction by Charnes et al. (1978). The domain of inquiry of DEA is a set of the entities, commonly called decision making units (DMUs), which receive multiple inputs and produce multiple outputs. Given a sample of the DMUs, the purpose of DEA is establishing of the relative efficiencies of each DMU within a sample. By collapsing the multiple inputs and outputs into the meta input and meta output, DEA employs linear programming techniques to establish an input–output based ratio that is then used to derive and assign a relative efficiency score to each DMU in the sample. As any non-parametric method, DEA eschews the hard-to-satisfy distributional assumption of data normality, thus allowing to investigator to inquire into the nature of the relative efficiency of the sample of convenience. This comes with the price, however, for DEA relies on a different set of assumptions that must be satisfied for the results to be valid. One of the fundamental assumptions of DEA is that of a functional similarity of the DMUs in a sample. This simply means that in order to compare, meaningfully, the relative efficiencies of DMUs in the sample (or data set), these DMUs must be similar in terms of utilization of the inputs and production of the outputs. Thus, DEA requires us to make sure that we compare apples and apples, and not apples and oranges. The difficulty arises when DMUs are represented not by apples and oranges, but by somewhat less homogenous entities such as people, companies, schools, and countries. 245

246

Quantitative Methodologies using Multi-Methods

In general, departments are more complex than people, companies are more complex than departments, and industries and countries are more complex than companies. As a complexity of a DMU increases, it gives rise to the increase in diversity in the set of DMUs, and this leads, problematically for DEA, to heterogeneity, rather than homogeneity, of the sample. One of our motivations for addressing is this issue is that it arises in various applications of the DEA methodology to real world problems. In our case we were exploring the relative efficiencies of a set of countries that were transitioning from centralized, planned economies to the market economies. To the extent that they are often considered by the international community, including significant actors such as the World Bank and IMF, to be a single group under the label Transition Economies (TE), there appears to be good reason to meaningfully compare them using DEA. However, in the situations where a domain of DMUs has been superimposed by the existing classifications, i.e., TEs, firms, departments, industries, one may end up with the sample consisting of DMUs that are functionally similar, yet heterogeneous. The DMUs are functionally similar because they receive, according to the specified DEA model, the same set of the inputs and produce the same set of the outputs, and such DMUs are possibly heterogeneous because no DMU could be specified completely by its inputs and the outputs. Again, complexity of the DMUs gives rise to the heterogeneity of the relevant set. It would appear that while the functional similarity of DMUs in a sample is assumed, the homogeneity of the DMUs is taken for granted. This, of course, severely limits the discriminatory power of DEA results. On the other hand that the explicit consideration of the possible heterogeneity in a sample would increase the discriminatory power of the results of the DEA. Currently, it would appear that the test of the assumption of the homogeneity of the DMUs in the sample represents the less rigorous, if not non-existent, part of DEA, for it is implicitly assumed that an investigator made sure that this important assumption holds. We suggest that the rigor of DEA, as well as its discriminatory power, could be increased by making the process of the assumption checking more explicit and objective. If one intends to inquire into the differences between efficient and inefficient DMUs in the sample, then, fundamentally, there are only two possible routes. First, if the assumption of the homogeneity holds, then the differences in the scores of the relative efficiency could be investigated without any adjustments. Or, second, if the assumption of the homogeneity of the DMUs in the sample does not hold, then the differences in the scores of the relative efficiency should be investigated with the appropriate adjustments. The question, then, becomes: What are the appropriate adjustments? It is not our intent to propose that DEA should only be performed in the situations where homogeneity of the DMUs has been decisively established. After all, a homogeneity, or, heterogeneity, of the DMUs in a sample is a matter of a degree. Sometimes the differences are minor and the researchers should declare that essentially the assumption of the homogeneity of the DMUs holds. Other times, however, the researchers might want to take into consideration the differences between the DMUs, while still comparing their relative efficiencies within the same sample.

Increasing the Discriminatory Power of DEA

247

Currently, to our knowledge, there exists no methodology allowing a researcher to investigate the differences between the relatively efficient and inefficient DMUs while taking into consideration heterogeneity of the DMUs in the sample. In this research we aim to address this problem. Namely, we propose and illustrate a threestep methodology allowing a researcher, first, to inquire into the differences between the DMUs in the sample, second, conduct DEA, and third, inquire into the differences between the relatively efficient and inefficient DMUs while taking into consideration the differences uncovered in the first step. The first step of our methodology utilizes cluster analysis (CA), the second step employs DEA, and the third step relies on the decision tree (DT) analysis. We provide brief overview of these data analytic techniques next, followed by the description of the proposed methodology.

THE PROPOSED METHODOLOGY In this section, we describe, in step-by-step fashion, the sequence of the procedures constituting the proposed methodology. Before describing our methodology, we provide an overview of the data set that is used in our running illustrative example.

Overview of Data Set of Illustrative Example To illustrate our methodology in action, we have chosen the following problem. Given a 10-year data set on 18 TEs, spanning a period from 1993 to 2002, we want to find out what are: 1. The differences in the relative efficiencies of these economies regarding their investments in telecoms, and 2. Some of the factors that contribute to the differences in the relative efficiencies. The data for this study were obtained from two sources. The first source was represented by the database of World Development Indicators, which is the World Bank’s comprehensive database on development data. The second source of the data was represented by the Yearbook of Statistics, which is published yearly by International Telecommunication Union (ITU). In our choice of variables, we were greatly restricted by the availability of the data. For example, while the development data of the World Bank’s database covers more than 600 indicators for 208 economies, data on many of the indicators relevant to our research were not available, or were available only for a few countries, or contained too few data points to be useful in statistical analysis. In terms of the length of the time series, we were restricted to the period from 1992 to 2002, data for which were provided by Yearbook of Statistics of ITU. In our choice of TEs we were guided by the intent to isolate a group of countries that started the process of transition in approximately the same time. As a result, we have decided to concentrate on the 25 countries of the former Soviet bloc.

248

Quantitative Methodologies using Multi-Methods

TABLE 26.1 List of Variables for DEA Models Role Input

Output

Sub-set of Variables GDP per capita (in current US $), Full-time telecommunication staff (% of total labor force), Annual telecom investment per telecom worker, Annual telecom investment (% of GDP in current US $), Annual telecom investment per capita, Annual telecom investment per worker Total telecom services revenue per telecom worker, Total telecom services revenue (% of GDP in current US $), Total telecom services revenue per worker, Total telecom services revenue per capita

Based on the availability of the data, the following 18 transitional economies out of 25 have been selected for this research: Albania, Armenia, Azerbaijan, Belarus, Bulgaria, Czech Republic, Estonia, Hungary, Kazakhstan, Kyrgyz Republic, Latvia, Lithuania, Moldova, Poland, Romania, Slovak Republic, Slovenia, and Ukraine. Despite the original intent, the data offered for 7 out of 25 TEs, namely, Tajikistan, Turkmenistan, Uzbekistan, Georgia, Macedonia, Russian Federation and Croatia, turned out to be insufficient to allow the inclusion of these economies in this study. For the DEA part of the methodology we have identified a model consisting of the input variables and output variables that are listed in Table 26.1.

Description of the Methodology Our methodology has the following three major steps that will be described in detail: 1. Determine the structural homogeneity status of the data set 2. Determine the relative efficiency status of each DMU 3. Describe the relative efficiency categories These steps require the initialization of relevant parameters including: • kMax: the maximum possible number of clusters that would be of interest to the decision-maker. This parameter is required in Step 1. • τOutlier: the threshold that is used to determine if a cluster is an outlier. This parameter is required in Step 1. • DMU_Goal: This could be “Input Orientation” or “Output Orientation”. This parameter is required in Step 2. • DMU_Criterion: This could be CRS, VRS, or NIRS. This parameter is required in Step 2.

Increasing the Discriminatory Power of DEA

249

Step 1: Determine the Structural Homogeneity Status of the Data Set Description of Step 1 Sub-Step 1a: a. Apply two-step approach to generate segmentations of sizes kMax through kMin. b. Set k = k Max. Sub-Step 1b: Examination the segmentation with k clusters. IF k > 1 and there is at least one cluster is that consists of less than τOutlier percent of the DMUs  THEN       Set k = k – 1;      Repeat Substep 1b  ELSE      Current Segmentation with k clusters provides the “natural” groupings of the DMUs;      Terminate Step 1. Justification of Step 1 The intended purpose of this CA Step 1 of our methodology is to investigate a “structural similarity” of the data set. Structural similarity of the DMUs reflects not the types, not the transformation of the inputs into outputs, but the levels of the inputs and outputs that DMUs receive and produce. Consequently, in the first step of our methodology we aim to determine whether all DMUs in the sample are similar in terms of the levels of the received inputs and the levels of the produced outputs. It could be suggested that the purpose of CA is to test the assumption of the homogeneity of the domain given the chosen DEA model. Consequently, homogeneity of the domain of the DMUs is always going to be relative to the given DEA model. Thus, the required pre-requisite to the first step in our methodology is that an investigator has identified a DEA model, i.e., a set of the inputs and a set of the outputs, which are going to be used in the Step 2. Once the model is determined, the actual data set that is going to be used to perform DEA is subjected to CA. During the first stage of CA we suggest to start, assuming that an investigator uses a software package allowing to conduct CA, with generating automatically a baseline clustering solution. Once it is done and the certain large number of the clusters has been generated, an investigator should evaluate the membership of each cluster. We suggest, as a rule of thumb, not to retain any cluster with the number of DMUs less than τOutlier percent of the total data set. By gradually decreasing the number of clusters over the iterations of CA, an investigator would get one of the two types of the CA solutions. First type is reflected by the presence of a single large cluster, containing, as a rule of thumb, (100 − τOutlier) percent or more of the DMUs, and one or two small clusters, with combined membership of (again as a rule of thumb) τOutlier percent or less of the sample. If this is the case, we suggest that an investigator should treat the sample of DMUs as homogenous. In the case of the second type of CA solution, however, an investigator ends up with two or more clusters with the membership of more than 10% of the sample in

250

Quantitative Methodologies using Multi-Methods

each cluster. A possible case in this scenario is when the final solution is presented in the form of two or three large clusters and a single small cluster (less than 10% of the sample). In this situation we suggest a visual inspection of the solution (the diagrams are provided by most of the software packages), or examination of the distances between the clusters; based on the results of the evaluation a small cluster should be combined with the closest large cluster. If the CA yields a second type of a solution, we suggest that the sample should be treated as heterogeneous and this heterogeneity should be reflected in labeling the resulting clusters as “Cluster1”, “Cluster2”, etc. Consequently, an investigator should document the membership of each cluster, by noting which DMUs belong to which cluster. After it is done, we proceed to the next step, DEA. Illustration of Step 1 For our illustration of this step we decided to use following eight variables to inquire into the homogeneity of our data set: 1. Total telecom services revenue (% of GDP in current US $), 2. Total telecom services revenue per capita (Current US $), 3. Total telecom services revenue per worker (Current US $), 4. Total telecom services revenue per telecom worker (Current US $), 5. Annual telecom investment per capita (Current US $), 6. Annual telecom investment (% of GDP in current US $), 7. Annual telecom investment per worker (Current US $), 8. Annual telecom investment per telecom worker (Current US $). The parameter τOutlier was set to 10%, kMax was set to 5 and kMin to 2. We used SAS Enterprise Miner (EM) to perform CA of the data set. The variables that we used are not measured on the same scale, so, prior to CA we transformed the data by standardizing the variables. We started our analysis by choosing “Automatic” setting, which did not require any input regarding the desired number of clusters from the researcher. Summary information regarding each obtained solution is compiled in Table 26.2. TABLE 26.2 Summary Output of Clustering Number of Clusters 5 clusters

Number of DMUs in Each Cluster 10, 11, 20, 3, 136

4 clusters

10, 32, 3, 135

3 clusters 2 clusters

30, 3, 147 72, 108

Top-level Split Variables Total telecom service revenue per worker Annual telecom investment per worker Annual telecom investment (% of GDP) Total telecom service revenue per worker Annual telecom investment per worker Total telecom services revenue per worker Annual telecom investment per telecom worker

Clusters with less than τOutlier percent of the DMUs are underlined.

251

Increasing the Discriminatory Power of DEA

TABLE 26.3 Cluster Membership when Number of Cluster = 2 Contents of the 1st Cluster Albania (1993–2002) Armenia (1993–2002) Azerbaijan (1993–2002) Belarus (1993–2002) Bulgaria (1993–2001) Slovak Rep (1993, 1994, 1999) Kazakhstan (1993–2002) Kyrgyz Rep (1993–2002) Latvia (1993, 1996) Lithuania (1993–1998) Moldova (1993–2002) Romania (1993–2002) Ukraine (1993–2001)

Contents of the 2nd Cluster Czech rep (1993–2002) Estonia (1994–2002) Hungary (1993–2002) Bulgaria (2002) Latvia (1994, 1995, 1997–2002) Lithuania (1999–2002) Slovenia (1993–2002) Poland (1993–2002) Slovak Rep (1995–1998, 2000–2002)

By using CA, we were able to come up with a solution that partitions our data set into two clusters. The membership of each cluster is provided in Table 26.3. Based on the compiled information we can see, that while some of the TEs are “permanent residents” of one cluster, other TEs are “migrants”, i.e., they change the cluster membership depending on a year. Finally, we should ask ourselves the following question: What is the significance of the separation of 18 transitional economies into the two clusters? One of the possible answers is provided in the research by Piatkowski (2003), who concluded that in the period “between 1995 and 2000 ICT capital has most potently contributed to output growth in the Czech Republic, Hungary, Poland, and Slovenia”. Thus, it could be suggested that we were able to separate 18 transitional economies into the two groups, one group of transitional economies that benefits the most from the investments in telecom, and another group where the benefits are less pronounced. Consequently, we labeled the members of the Cluster 1 as “Majority” and the members of the Cluster 2 as “Leaders”. Step 2: Determine the Relative Efficiency Status of DMUs Description of Step 2 Based on the values of the parameters DMU_Goal (i.e., “Input Orientation” or “Output Orientation”) and DMU_Criterion (i.e., CRS, VRS, or NIRS), the relevant DEA approach is performed in order to obtain the relative efficiency scores of the DMUs in the sample. We assume that at the end of the DEA process each object is assigned a specific relative efficiency status (relatively efficient, or relatively inefficient), and so each object has an additional relative efficiency status attribute (say EfficiencyStatus with the value “1” for Relatively Efficient, and the value “0” for Relatively Inefficient).

252

Quantitative Methodologies using Multi-Methods

Illustration of Step 2 For our illustration of this step, given the values of the parameters DMU_Goal and DMU_Criterion, the relevant DEA models need to be constructed. Although not required, we generated the relative efficiencies of DMUs using both input-oriented and output-oriented DEA models, for three types of conditions, constant (CRS), variable (VRS), and non-increasing return to scale (NIRS). Thus, we generated altogether six DEA models, with the following settings of the parameters DMU_Goal and DMU_Criterion: 1. DMU_Goal = “Input Orientation” and DMU_Criterion = CRS 2. DMU_Goal = “Input Orientation” and DMU_Criterion = VRS 3. DMU_Goal = “Input Orientation” and DMU_Criterion = NIRS 4. DMU_Goal = “Output Orientation” and DMU_Criterion = CRS 5. DMU_Goal = “Output Orientation” and DMU_Criterion = VRS 6. DMU_Goal = “Output Orientation” and DMU_Criterion = NIRS This step does not differ in any way or form from the regular DEA, thus, no adjustments are necessary. We also computed the average relative efficiencies of the two clusters identified in the Step 1, and compare the averaged relative efficiencies of these two groups of TEs produced by the DEA. Table 26.4 demonstrates the differences in the relative efficiency between the “Leaders” and the “Majority” in terms of the utilization of the inputs. The input-orientation does not concern itself with the maximization of the outputs, but rather with maximization of the utilization of the inputs. Thus, it is probably reflective of the perspective of the policy maker, especially in the case when the available resources are limited. Table 26.5 demonstrates the differences in the relative efficiency between the “Leaders” and the “Majority” in terms of the maximization of the outputs. Unlike in the case of the input-oriented model, the output-orientation does not concern itself with the efficient utilization of the inputs, but rather with the maximization of the outputs. Thus, it is probably reflective of the perspective of the investor, especially in the case when the resources are abundant and the primary goal is to obtain the maximum revenue.

TABLE 26.4 DEA: Comparison of the Clusters based on the Input-Oriented Model Criterion for Comparison Average efficiency score, CRS Average efficiency score, VRS Average efficiency score, NIRS Average efficiency score, SE

“Leaders” Cluster 0.89 0.95 0.89 0.94

“Majority” Cluster 0.79 0.88 0.80 0.89

Difference 0.10 0.07 0.09 0.04

Difference % 12.54%  7.48% 11.63%  4.96%

253

Increasing the Discriminatory Power of DEA

TABLE 26.5 Comparison of the Clusters based on the Output-Oriented DEA Model Criterion for Comparison Average efficiency score, CRS Average efficiency score, VRS Average efficiency score, NIRS Average efficiency score, SE

“Leaders” Cluster 1.17 1.15 1.16 1.02

“Majority” Cluster 1.41 1.29 1.36 1.10

Difference −0.24 −0.14 −0.19 −0.07

Difference % −16.71% −11.00% −14.29% −6.81%

Out of the six DEA models generated in the Step 2 we have arbitrarily chosen the model DMU_Goal = “Input Oriented” and DMU_Criterion = “CRS” to illustrate Step 3 of our methodology. Step 3: Describe the Relative Efficiency Categories Description of Step 3 In this step, we will use DT induction to generate rules that can describe the relative efficiency categories in terms of the input and output variables of the DEA models. This will require the inclusion of a target variable (say “EfficiencyCategory”) that identifies the efficiency category. For the case of the homogeneous sample of DMUs, we would use a binary variable to indicate whether the DMU is relatively efficient. For a homogenous data set, we have to use a categorical variable that allows for two efficiency categories per cluster. Thus, in the case if the CA in the Step 1 resulted in the solution with two clusters, the domain of values for our categorical target variable could be represented as follows: “11” – Relatively Efficient DMU with membership in the Cluster 1, “10” – Relatively Inefficient DMU with membership in the Cluster 1, “21” – Relatively Efficient DMU with membership in the Cluster 2, “20” – Relatively Inefficient DMU with membership in the Cluster 2. If we do not take into consideration heterogeneity of the sample, then we end up with the following domain of values: “1” – Relatively Efficient DMU in the sample, “2” – Relatively Inefficient DMU in the sample. In general, we will assume that our target variable (say EfficiencyCategory) is a concatenation of the relevant cluster identifier attribute ClusterNum and relative efficiency status attribute EfficiencyStatus. Once the data set has been amended to include the target variable EfficiencyCategory, DT induction is used to generate a DT that can be used to describe the efficiency categories.

254

Quantitative Methodologies using Multi-Methods

Illustration of Step 3 For our illustration of this step, since the result of Step 1 indicated that our data set was heterogeneous with two groups, we first populated our target variable “Cluster Efficiency” with the following values: • Value of “21” was assigned to the “Efficient Leaders”, those TEs that belong to the “leaders” cluster and were assigned the score of “1” by DEA • Value of “20” was assigned to the “Inefficient Leaders”, those TEs that belong to the “leaders” cluster and were assigned the score of less than “1” by DEA • Value of “11” was assigned to the “efficient majority”, those TEs that belong to the “majority” cluster and were assigned the score of “1” by DEA • Value of “10” was assigned to the “Inefficient Majority”, those TEs that belong to the “Majority” cluster and were assigned the score of less than “1” by DEA The number (N) of DMUs in each of our four categories are: 38 Efficient Leaders, 31 Inefficient Leaders, 43 Efficient Majority, and 68 Inefficient Majority. We then generated our DT model that enabled us to identify conditions associated with our four categories. In Table 26.6 we display a sub-set of the rules that are associated with this DT. For each category, we selected a pair of rules, each of which had a strong probability (i.e., Prob > 0.90) for the occurrence of the associated category given the condition component of the rule.

TABLE 26.6 Pairs of Rules that Describe the Efficiency Categories Efficiency Category Condition Productivity Ratio per Telecom Worker ≥ 4.1754445351 & Annual Telecom Investment per Worker ≥ $58 Total Telecom Services Revenue per person ≥ $210 & Full-Time Telecommunication Staff % ≥ 0.0039016912 & Productivity Ratio per Telecom Worker < 4.1754445351 & Annual Telecom Investment per Worker ≥ $58 Full-Time Telecommunication Staff % < 0.0039016912 & Productivity Ratio per Telecom Worker < 4.1754445351 & Annual Telecom Investment per Worker ≥ $58 Full-Time Telecommunication Staff % < 0.0031414015 & Productivity Ratio per Telecom Worker ≥ 3.8043909395 & GDP per Capita ≥ $519 & Annual Telecom Investment per Worker < $33

Group Leader

Efficiency Status Efficient

N 14

Prob 1.00

Leader

Inefficient

 5

1.00

Leader

Inefficient

11

1.00

Majority

Efficient

 8

1.00

255

Increasing the Discriminatory Power of DEA Efficiency Category Condition Total Telecom Services Revenue ≥ 0.0118204323 & GDP per Capita < $519 & Annual Telecom Investment per Worker < $33 Productivity Ratio per Telecom Worker < 3.8043909395 & GDP per Capita ≥ $519 & Annual Telecom Investment per Worker < $33 Full-Time Telecommunication Staff % ≥ 0.0031414015 & Full-Time Telecommunication Staff %< 0.0054371357 & Productivity Ratio per Telecom Worker ≤3.8043909395 & GDP per Capita < $519 & Annual Telecom Investment per Worker < $33 Productivity Ratio per Telecom Worker < 2.002357802 & $33 ≤ Annual Telecom Investment per Worker < $58

Group Majority

Efficiency Status Efficient

N 22

Prob 1.00

Majority

Inefficient

39

1.00

Majority

Inefficient

12

0.92

Majority

Inefficient

 6

1.00

CONCLUSION DEA is a good data analytic tool discriminatory power of which, however, is somewhat dependent on the homogeneity of the sample of DMUs. Relative efficiencies of the NFL players, for example, could not be meaningfully compared unless the position played, which could be dependent on height, speed, and weight, is taken into consideration. Similarly, relative efficiencies of the designated hitters and catchers of MLB in terms of the production of the home runs should not be compared directly either. Thus, necessity arises to find a way to conduct DEA while taking into consideration some of the important differences between the groups of the DMUs in the sample. In the case of this study, we have identified two sub groups within our sample of 18 TEs, the “Majority” and the “Leaders”. Having this insight in mind, we have three options of conducting DEA. First, we could disregard this information and proceed directly with DEA. Second, we may conduct separate DEA per each cluster. Third, we have an option of proceeding according to the proposed in this chapter methodology. We would like to argue that our methodology allows for achieving of a better discriminatory power of DEA than its two alternatives. In order to do so, we offer a comparison of the pairs of rules that were obtained by utilizing each of the three mentioned above options. In order to avoid repeating contents of Table 26.6 in this section of the chapter, we provide only the sets of the decision rules corresponding to two other options. 1st option: DT based on DEA only (“efficiency” encoded as “1” – efficient, “0” – inefficient), 57 efficient DMUs, 123 inefficient DMUs The results of this DEA demonstrate that a number of transitional economies have obtained a rating of being relatively efficient. It does not mean, however, that all of the countries that were deemed relatively efficient are in fact efficient. A common

256

Quantitative Methodologies using Multi-Methods

characteristic of DEA models is that they tend to evaluate as efficient those DMUs that have the smallest input values, or, the DMUs with the largest outputs (Ali, 1994). Thus, the approach consistent with the 1st option does not allow an investigator to determine whether the relative efficiency of a DMU is caused indeed by its efficiency, or whether a DMU was awarded a relatively efficient status because it had the smallest level of the inputs in the sample. Consequently, one of the shortcomings of the first DT model is that it does not allow an investigator to incorporate the results of the CA, i.e., an “Efficient Leader” is highly likely to be very different from an “Efficient Majority”. 2nd option: DT based on DEA, one DT model per cluster a. “Leaders” cluster (“efficiency” encoded as “1” – efficient, “0” – inefficient), 38 efficient DMUs, 31 inefficient DMUs b. “Majority” cluster (“efficiency” encoded as “1” – efficient, “0” – inefficient), 43 efficient DMUs, 68 inefficient DMUs These models, consistent with the 2nd option, improve on the first model by being cluster-specific. Let us recall, nevertheless, that while some of the TEs are “permanent residents” of one cluster, the other TEs are “migrants”, for they change their membership depending on the year. Consequently, none of the generated by this approach models would help us to inquire into the question why, for example, Lithuania was a member of the “Majority” cluster for the period from 1993 to 1998, but became a member of the “Leaders” in the period from 1999 to 2002. Considering the presented above alternatives, it would appear that the results of our methodology, presented in Table 26.6, allow for achieving of a higher discriminatory power of DEA. While the results of the conventional DEA only yield the efficiency scores for each DMU in the sample, the results of our approach yield the efficiency scores and, in the case of the heterogeneous domain, the membership within the subset of the sample for each DMU. This (considering that the most decision makers are interested not so much in learning whether or not a given DMU is inefficient, but rather why it is inefficient) allows for a higher degree of granularity of the subsequent analysis, as it was demonstrated by comparing Tables 26.6, 26.7, 26.8, and 26.9. Another benefit of our methodology is that it takes an approach of an “external augmentation” of DEA, meaning, it does not require any changes to or alterations of DEA itself. As a result, our methodology is not model-specific and, consequently, could be applied to any DEA model. However, despite the contributions that this research makes, we must acknowledge that our study is not without its limitations. First limitation of this research is associated with the use of CA. At this point, we cannot offer strict criteria determining whether the sample should be considered homogeneous or heterogeneous. Thus, despite the explicit nature of testing for heterogeneity by means of CA, the determining decision regarding the sample still lies with the decision maker. Second limitation of our study is associated with the use of DT induction. At this point, we cannot suggest to a decision maker what splitting criteria and what settings yield a better tree. As a result, this issue as well resides in the domain of the responsibilities of the decision maker.

257

Increasing the Discriminatory Power of DEA

TABLE 26.7 First Option: Pairs of Rules that Describe the Efficiency Categories Efficiency Category Condition Full-time telecommunication staff % < 0.0041924905 & Productivity ratio per telecom worker ≤ 5.4191734889 Total telecom services revenue ≤ $51,715,239 & Total telecom services revenue per telecom worker < $60,794 & Productivity ratio per telecom worker < 5.4191734889

Group Whole Set

Efficiency Status Inefficient

N 25

Prob 0.92

Whole Set

Inefficient

91

0.956

TABLE 26.8 Second Option: Pairs of Rules that Describe the Efficiency Categories in “Leaders” Cluster Efficiency Category Condition Total telecom service revenue per capita ≤$101 & Productivity ratio per telecom worker < 4.1754445351 & Annual telecom investment % GDP < 0.0166 & Full-time telecommunication staff % ≤0.0043712305 Productivity ratio per telecom worker ≥ 4.1754445351 Full-time telecommunication staff % < 0.0043712305 & Productivity ratio per telecom worker < 4.1754445351

Group Leaders

Efficiency Status Efficient

N 12

Prob 1.0

Leaders Leaders

Efficient Inefficient

17 21

1.0 0.905

TABLE 26.9 Second Option: Pairs of Rules that Describe the Efficiency Categories in “Majority” Cluster Efficiency Category Condition Productivity ratio per telecom worker < 3.8106041166 & GDP per capita ≤ $519 Total telecom services revenue % ≤0.0118204323 & GDP per capita < $519

Group Majority

Efficiency Status Inefficient

N 47

Prob 1.0

Majority

Efficient

22

1.0

258

Quantitative Methodologies using Multi-Methods

ACKNOWLEDGMENT Material in this chapter previously appeared in: Increasing the discriminatory power of DEA in the presence of the sample heterogeneity with cluster analysis and decision trees. Expert Systems with Applications, 34(2), 1568–1581.

REFERENCES Ali, A.I. (1994). “Computational aspects of DEA.” In: Charnes, A., Cooper, W.W., Lewin, A., and Seiford, L.M. (eds.), Data Envelopment Analysis: Theory, Methodology and Applications. Kluwer Academic Publishers, Boston, MA, pp. 63–88. Charnes, A., Cooper, W.W., & Rhodes, E. (1978). Measuring the efficiency of decision making units. European Journal of Operational Research, 2, 429–444. Charnes, A., Cooper, W.W., Lewin, A.Y., and Seiford, L.M. (1994). Data Envelopment Analysis: Theory, Methodology and Applications. Kluwer Academic Publishers, Norwell, MA. Hong, H., Ha, S., Shin, C., Park, S., and Kim, S. (1999). “Evaluating the efficiency of system integration projects using data envelopment analysis (DEA) and machine learning”, Expert Systems with Applications, Vol. 16, pp. 283–296. Seol, H., Choi, J., Park, G., and Park, Y. (2007). “A framework for benchmarking service process using data envelopment analysis and decision tree”, Expert Systems with Applications, Vol. 32, No. 2, pp. 432–440. Sohn, S., and Moon, T. (2004). “Decision tree based on data envelopment analysis for effective technology commercialization”, Expert Systems with Applications, Vol. 26, No. 2, pp. 279–284.

27

An Exploration of the Intrinsic Negative SocioEconomic Implications of ICT Interventions

INTRODUCTION The socio-economic impacts of information and communication technology (ICT) have been extensively researched in diverse contexts, within different time frames, and by various methods of inquiry. It is fair to say there is a general consensus that investments in ICT result in “good things” of various kinds. After reading findings of published research, one feels that investing in ICT is similar to buying a nighttime cough medicine to help your cold feel better next morning. You pay the price (for ICT does not come free), you take the medicine (ICT needs to be integrated and implemented), and you wait (there is a time lag). And it usually works as advertised – you do feel better and all seems to be well. The medicine does not work perfectly, but it never completely fails. So, overall, this medicine, as well as ICT, is a “good thing”. However, a medicine always comes with a warning of side effects – there is a disclosure mentioning an additional price to pay for getting a “good thing” going. Unlike medicine, side effects of “good things” brought about by ICT are almost impolite to mention – for the technological progress and productivity growth usually come first and rank high on any economy’s list of priorities. And they should come first, for it is better to make money, generate wealth, and then re-distribute them if needed, than to be ready to re-distribute, but having nothing to give. But there is a price to pay for economic development – side effects are there whether we want it or not. Technology and progress are like a big ship taking off the pier and heading full speed to the sea – moving economies forward, but leaving some small boats greatly disturbed in its wake. We should not try to stop, or slow down, this big ship, but we should be aware of what it leaves in its wake, so we are prepared for little boats to be in trouble. This is, really, a matter of comparing and contrasting “big deals”. Increasing efficiency and effectiveness of a business process, or a governmental service, or an Internet access is a big deal, indeed. We can trace the impact to growth in productivity, and then, in turn, to the macro-economic outcome – growth in GDP. But there are other “bid deals”. Big deal if a local mom-and-pop store went out of business – are you not happy, instead, to use a self-checkout station in a local supermarket? Well, a cashier that 259

260

Quantitative Methodologies using Multi-Methods

knew your name is gone, but the process is much quicker now, no cashier is needed and prices are lower. Much lower than in that mom-and-pop store, now out of business, where locals used to meet. Some of the negatives are situation-specific and unique to a given setting offering researchers an anecdotal evidence. Other negatives are philosophically inevitable – they will be present regardless of the context. If you are in business of designing bombs, then you should aim for making the most efficient and effective once – just business, nothing personal. However, it is good to be aware of a collateral damage of a bomb maker’ success. Our objective in this editorial is to bring attention of our colleagues to some of the “negative inevitables” of positive impacts of ICT (Bosamia, 2013; Tarafdar et al., 2013, 2015). Along the way we will try to suggest a way of categorizing the negatives. Finally, we will attempt to propose possible routes of investigating them. A full disclosure to our readers – it is not our purpose to rain on a parade of positive impacts of ICT, but it is our goal to suggest to our colleagues that ICT-driven initiatives aiming for “good things” should come with a warning label mentioning side effects.

SOCIO-ECONOMIC IMPACT OF ICT Tools, Machines, and ICT Before ICT was a “good thing”, there were machines, and before machines there were tools – “good things”, all. It is important to mention them because they are inextricably linked together. An introduction of tools, machines, and ICT has a shared objective – to make an execution of work more effective and efficient, which pursued according to the same strategy, that of elimination and substitution. But while tools, machines, and ICT share a common objective and a common strategy, they implement them by following different routes. Let us spend a minute clarifying terminology used in this chapter. Physical work is an act of applying energy to matter with the purpose of transforming its properties. Physical work has its counterpart in digital domain. Digital work is as an act of collecting, manipulating, storing, or disseminating of data – an act of applying energy to data with the purpose of changing its properties – location, format, encoding, encryption, etc. A work element is a part of work, a task, that could be performed by a single worker. Sharpening of a knife or an entry of a record into a database are examples of work elements. A tool is an object, natural or manmade, allowing for a more efficient and effective performance of a function within a work element. A knife sharpener and a document scanner are examples of tools. A machine is a system incorporating a tool(s) allowing for a more efficient and effective performance of work. A knife sharpening machine is an example of a machine in the mechanical domain. A digital graphing calculator is an example of a digital machine. An Information Technology (IT) is an electronic symbol generating and processing system comprising software and hardware that allows for assessing, via

Intrinsic Negative Socio-Economic Implications

261

generating information, the working order, or a state, of a machine. IT-enabled sharpening machine allows for monitoring a number of surfaces sharpened, a wearing out of the sharpening tool, and the time-to-service left – once it is up the warning message appears. A laptop is an example of IT that controls the state of multiple virtual machines – applications. ICT is an IT-based system that creates new, or optimizes existing, information channels which allow for controlling a set of virtual or IT-controlled machines involved in a complex workflow (work processes). An automated order fulfillment center and a cloud-based online payment system are examples of ICT.

Routes of Elimination and Substitution Let us take a look at various ways by which tools, machines, and ICT bring about “good things”. A tool allows for increasing efficiency and effectiveness of performing a work element via eliminating a direct involvement of a human body part(s) in performing work by means of substituting it with an object. Hitting your enemy in the head using a hammer is more efficient and effective than using a fist. Digitally recording a song is more effective and efficient than doing so in analog way, whether recording it on a vinyl disk or via transcribing its text and notes. A machine allows for increasing efficiency and effectiveness of work via mechanization or automation – by eliminating human involvement in repetitive work by means of substituting humans equipped with tools. IT allows for increasing efficiency and effectiveness of work via substituting humans controlling the machines. An automated production line’ control system, or a self-checkout counter at the grocery store both eliminate significant human involvement. ICT takes it a step further and achieves the objective via substituting human decision-making process geared toward an optimization of a workflow. Just-in-time (JIT) manufacturing system or Apple’ iTunes store are examples here. In both cases human involvement is minimized. An artificial intelligence (AI)-enabled ICT aims to replace human decision making within a novel context (e.g., changed business environment) of a workflow. Collaterally, AI is a tool that attempts to manipulate human decision-making processes. The impact of elimination and substitution via tools, machines, and ICT is compound because the elements in the elimination and substitution chain are interdependent. Better tools allow for integrating them in more complex machines that are controlled by the increasingly sophisticated software component of IT, and the complex machines could be interconnected via ICT (especially AI-enabled ICT) to comprise a complex workflow. Consequently, if entity A invests in a more sophisticated ICT than entity B, then, ceteris paribus, entity A should be able achieve a greater level of effectiveness and efficiency of work via elimination and substitution alone. However, there is inevitable impedance mismatch within the elimination and substitution chain if the chain incorporates analog physical work and digital symbolic processing and control. This means that even the most sophisticated warehouse or distribution center (e.g., Amazon’ ones) will not be able to eliminate as many people

262

Quantitative Methodologies using Multi-Methods

as would be possible otherwise. The maximum effect the chain achieves in the case of moving all the work into a digital domain (e.g., Instagram replaced 100K Kodak’ employees). For example, a branch of a bank with an on-ground presence will have to have some on-ground employees left as long as they deal with physical money. If, however, the bank or a financial institution moves its operations completely into a digital domain, then the success of the substitution and elimination chain could be much greater. In any case, digital or analog, any implementation of ICT will result in inevitable elimination and substitution, somewhere along the chain, of human involvement. Furthermore, this impact remains regardless the type of workflow – whether the work is accomplished for the benefits of shareholders (e.g., any implementation of Enterprise Systems), or if it is done for personal, leisurely purposes (e.g., subscribing to a new service or downloading a new app to your smartphone). At this point we are in a good position to state Negative Inevitable #1: Investments in ICT will result in a substitution and elimination of human involvement.

Conditions for Elimination and Substitution In principle, there are two conditions when technology (e.g., tools, machines, ICT) will not be used to substitute and eliminate human involvement. First, when there is no available technology to do so and, second, when technology is available, but it is cheaper to use humans. There are some important implications resulting from these conditions. In the first case it is worth considering a situation under which “available technology” is equivalent to “affordable technology”, for both cases will result in the same outcome – there will be no substitution. In the instance of ICT this is relevant, for a competitive advantage nowadays comes not from an ability to do work (which is a given), but from an ability to coordinate complex workflows. And it is in the domain of coordination of work that ICT shines. Thus, by improving the chain “resource → tools → machines → IT-enabled machines” an economy that cannot afford ICT may only go so far. This means that an economy that cannot afford needed ICT will not be able to make more money to afford it, but will have to wait until prices of such ICT are reduced sufficiently to be affordable. Basically, dragging further and further behind those who can afford it now. In the second case, if we ask ourselves a question “Where is it cheaper to use humans than technology in 21st century?” we can come up with a surprisingly long list of places. In all of those contexts’ investments in ICT, however important, should not be considered an overriding priority. Instead, the priority should be in transitioning from human labor to machines, and then to IT-enabled machines. But, for many economies that cannot afford ICT there are external sources that could help with the funding. So, what would be a negative consequence of getting a technology that an economy cannot really afford or could do without because human labor is cheaper? To answer this question, we need to consider the complete chain “Resources → Tools → Machines → IT-enabled Machines → ICT” to see that the

Intrinsic Negative Socio-Economic Implications

263

introduction of ICT will expose shortcomings and conflicts within the next link below – “IT-enabled machines”. Basically, an introduction of advanced (for a given context) ICT will result in a situation of “OK, now we have the technology to connect IT-controlled machines, but how do we get the machines up to speed now?” At this point we can state Negative Inevitable #2: Introduction of advanced-for-the-context ICT will magnify the existing, and identify new, deficiencies within the state of IT-enabled machines.

Pragmatics and Ethics of Implementation A different set of potential problems emerges if we consider the inevitable gap between an output of the implementation of ICT and the outcome of it. Investments in ICT are allocated primarily towards impacting the current infrastructure of the context and, naturally, the intended result is delivered via the implemented ICT, which is not an end to itself, but a mean to a greater end – a positive socio-economic impact. Overall, the following chain of links is traceable: “Investments in ICT → Implementation of ICT → Impact of Implemented ICT”, which is, for all intents and purposes, a chain “Investments in ICT → Output → Outcome”. Thus, the actual resultant ICT is the output of the process of implementation of investments in ICT. It is of no use to have investments in ICT available as an end game, for it is just a line on a balance sheet that makes an account look good, but doing not much beyond that. Consequently, investments must be allocated and implemented to produce an impact, and the most common way to introduce new technologies is via using ISD methodologies. The overwhelming majority of methodologies used in the process are functionalist – based on mechanistic approach and driven by pragmatic considerations. The quality of the output, therefore, is judged based on the appropriate to functionalism criteria, which, in the case of the chain “Resource → Tool → Machine → IT-controlled Machine → ICT”, translates into increased effectiveness and efficiency of the chain. A greater impact of ICT, however, resides not in the output of the process of implementation, but in the outcome of the process within a larger system. This forces us to move out of a narrow technological domain into a more complex social domain, where the assessment of the impact of ICT is no longer driven by pragmatics, but, instead, by ethics. Resultantly, it is only expected to anticipate a presence of incongruence between the output of ICT and the outcome of it, for pragmatics and ethics are based on different principles and assumptions. Now we are well-equipped to state Negative Inevitable #3: An assessment of the impact of ICT will require reconciliation of differences between the pragmatics of the implementation of ICT and the ethics of the expected results of the implementation. Let us spend some time considering the aspects of a social domain that could be impacted, positively and negatively, by ICT.

264

Quantitative Methodologies using Multi-Methods

Dimensions of Social Impact of ICT Introduction of ICT results in changes in the host’s social environment – even if the implementation itself is purely technical, impacts of the implementation spread beyond the boundaries of purely technological domain. Regardless of the context, no society is homogeneous, and we need to consider what sort of dimensions are responsible for heterogeneity, for every one of them will be impacted by ICT. We will consider only the most obvious ones – economic, cultural, political, and religious. At this point we are not going to be concerned with the “interaction effects” such as “wealthy liberal”, “poor conservative”, and so on. An uber dimension that is responsible for differentiating a society is economics. In the case of introducing a new and improved ICT the financially better off groups of society will disproportionally benefit from the change. After all, it is the members of any society that have more advanced IT that benefit from ICT the most. Consider having a new high-speed highway – the introduction of a new road is a good thing for any driver, but the drivers having better cars will benefit the most, while the owners of the old ones may struggle to adjust their vehicles to new conditions (e.g., higher speed limit). At this point Negative Inevitable #4 seems to be in order: Economically disadvantaged groups of a society will be negatively impacted by the introduction of new ICTs due to the pressures of the compliance with the new technology and the limited use of new capabilities due to restrictions associated with the old technology. However, pressures on a technological front are only a beginning of the impact of new ICT – let us consider other ones as well.

Platform, Message, and Target It is fair to say that almost any coherent, organized, and recognizable/identifiable sub-group in a society has its own platform and its own message, as well as an intended target of the message. We define a platform as a collection of the available information channels allowing for transmitting of the group’s message. A message is a statement containing a set of propositions reflecting the political, cultural, or religious values (or a combination of them) that the group wants to impart to a target. A target is an intended audience of the message, which could be another sub-group, or a society in general. Introduction of new ICT brings some interesting consequences impacting platforms, messages, and targets that are worth considering. A new ICT will allow for creating new and optimizing existing information channels. Thus, in the case of an affluent sub-group, its platform will become wider due to quantitative and qualitative changes the new ICT will bring. Quantitative changes are associated with the platform gaining a greater number of information channels of the same type, and qualitative changes are associated with gaining a greater variety of the types of information channels (e.g., text, audio, video, and digital art). The platform will also become taller via allowing for reaching a wider audience – a bigger target group. So, a new ICT is a “good thing” for a well-off group.

Intrinsic Negative Socio-Economic Implications

265

In the case of an economically disadvantaged sub-group, however, the platform will become relatively narrower, because it will require a newer technology to utilize new ICT to the full extent. Also, the platform will become relatively lower, because some of the previously accessible audiences will transition to using new technology that is not available to the poorer sub-group. It is possible, of course, that the new ICT will not result in reduction of the width and height of the platform of an economically disadvantaged group. However, because the economically advantaged group will take an opportunity of the new ICT, this will make their platform wider and taller than it previously was. Hence our use of the term relatively – even if the platform of the disadvantaged group stays the same, its width and height deteriorates relative to the competition. Consequently, Negative Inevitable #5 is as follows: Introduction of new ICT will result in a relative deterioration of the social platform of the economically disadvantaged subgroups of a society. Some interesting things happen to messages. While the substance, the content of the message may stay the same (it does not have to, but let us assume the simplest scenario), the two characteristics of a message will be impacted greatly. These characteristics are format and intensity. New ICTs offers new types of information channels, which allow for a delivery of a message in a greater variety of formats (e.g., text, audio, and video) and with a greater frequency. Increasing a variety of formats and a frequency of delivery neither come free, nor cheap. Hence, an introduction of new ICT will result in a relative decrease of intensity and a relative decrease of attractiveness of the format of the message in the case of economically disadvantaged sub-groups. It is important to note that we are not talking about a content – the substance of the message, but only about its’ packaging and a frequency of delivery. Resultantly, we put forward Negative Inevitable #6 as follows: For economically disadvantaged subgroups an introduction of new ICT will bring a relative decrease in (1) frequency of delivery and (2) attractiveness of the format of their message. Introduction of a new ICT also impacts targets. It is safe to say that the majority of active sub-groups in any society aim to reach outside its own constituency to promote their set of values. Let us consider a simple scenario where a sub-group has its loyal following and constituency that will not be impacted by the introduction of new ICT. The outside targets, however, will be impacted by the new ICT – the impact is based on what we call continuously rising expectation regarding the wrap and drop. It is only to be expected that people expect messages to be nicely wrapped – delivered to them in the best possible/available format (e.g., high-production video vs. text), and continuously dropped – delivered to them regularly and often. Targets, whether they want it or not, will become fickler as they are offered more choices which raise their expectations.

266

Quantitative Methodologies using Multi-Methods

It is only reasonable to suggest, then, that format wars and delivery battles will take place, and will play out with fairly predictable results. Economically better off sub-groups will win format wars because they have a financial edge in producing, or paying to produce, messages in a better wrapper – in a higher quality format. Also, it is only to be expected that more affluent sub-groups will win delivery battles because they will have more resources for delivering messages more frequently and on-demand – a fast food delivery model will prevail in digital domain. Simply put, a Darwinian’ message selection by a target plays out, where fast and pretty survives, and slow and substantive dies. Now we can state Negative Inevitable #7, as follows: For economically disadvantaged subgroups an introduction of new ICTs will make reaching the target of their message relatively harder. Let us now consider some political and cultural implications of ICT, where two negatives come to mind. First, an introduction of an advanced ICT will result in a further consolidation of political powers around the affluent centers. The reason is simple – the fundamental architecture of the Internet is based on a client-server model, and the centers of control and influence are usually formed around the servers. Consequently, if the resources of an economy are in the hands of a government (or, government – controlling entities), then mechanisms of political suppression will become more effective and efficient, specifically in the context of non-democratic societies. Simply put, bigger, better, faster servers will force re-centralization of decentralized model of the Internet-based computation, with all the political impacts it may bring. We suggest a Negative Inevitable #8, as follows: In the societies of a globalized world introduction of new ICT will strengthen the existing political status quo. A second implication is associated with the purpose of the culture of a society, which is to provide a shared, stable and consistent world-view to its members. An introduction of such change agent as advanced ICT may not result in a significant cultural impact in the context of the Western world, but in a context of societies that are based on different cultural premises it may. We should not forget that ICT is a messenger of such powerful mechanistic ideas as positivism, determinism, and individualism, and it requires a compliance with them to be utilized effectively and efficiently. It is not entirely unlike the democratic ideas of the West, which were successfully adopted by many post-Soviet countries in Europe, but did not root well in Arab world and Africa due to fundamental cultural differences. Simply put, ICT requires an appropriate cultural soil to take off, for, otherwise, cultural clashes are inevitable. We offer a Negative Inevitable #9, as follows: An expected corollary to the outcome of the implementation of the ICT in the non-Western context of the globalized world is a struggle to reconcile the values of the Western culture with the local one.

Intrinsic Negative Socio-Economic Implications

267

Additionally, there are also ICT-driven changes that impact inter-economy situation. We consider them next.

Competing with Others: Additional Implications An introduction of the new ICT will also have an inter-economy impact, for changes brought about by ICT rarely can be contained within a single economy. Let us consider two simple scenarios to illustrate the implications. The first scenario refers to the situation when Economy A wants to gain a competitive edge over Economy B via the introduction of the new ICT. We consider such situation playing out via two intuitive cases. The first case is when both economies have well-integrated and developed chains “Resources → Tools → Machines → IT-enabled Machines” and investing in new ICT does make a competitive sense for Economy A. However, the resultant competitive edge will be short-lived, for all it takes for Economy B to catch up is to invest in its own new ICT as well. Thus, once Economy B also gets its new ICT, the obvious next step for Economy A is to make the complete chain “Resources → Tools → Machines → IT-enabled Machines → ICT” leaner relative to Economy B. And this “leaning out” of a workflow chain will impact the involved labor force – elimination and substitution of the workforce will take place. However, the impact on the labor force will be greater because it will be driven by two factors working in synergy. The first factor is new ICT and its impact on a workflow chain, where “as-is” state with X employees will be replaced by “to-be” state with X-n employees. The second, compounding factor, is a competition with Economy B. Given the fact that Economy B will go through its own “as-is” to “to-be” transformation resulting in Y-m employees, Economy A will attempt to gain a competitive advantage via the second round of elimination and substitution aiming at prevailing over Economy B via assuring that the equation (X-n) < (Y-m) remains valid. The second scenario refers to the situation when Economy A, in the absence of the well-integrated and developed chain “Resources → Tools → Machines → IT-enabled Machines”, invests in new ICT aiming to attract outside investments (e.g., FDI) and, resultantly, gain an edge over Economy B via this route. However, it is only reasonable to assume that such plan will fail – two reasons come to mind. First, no external funds may come about if the outside investments will aim to take advantage of welldeveloped and integrated chain “Resources → Tools → Machines → IT-enabled Machines → ICT”. Second, external funds could be allocated under condition of improving the “Resources → Tools → Machines → IT-enabled Machines → ICT” chain. Because investments more often than not are allocated for the purpose of making profit, external entities will have a goal of using Economy A to compete against Economy X. This results in a familiar scenario of elimination and substitution resulting, first, in X-n as a result of new ICT, and second, (X-n)-k as a result of the pressure from external investors competing against Economy X. At this point we put forward Negative Inevitable #8, as follows: An introduction of new ICT as a tool of inter-economy competition will result in an increased level of elimination and substitution of the workforce.

268

Quantitative Methodologies using Multi-Methods

Competing with Others: Social Implications Inter-economy competition would also bring about some interesting political, cultural, and religious implications. While the extent of the impact along those dimensions is context-specific and indicative of local inflections, the general tendency can be traced. ICT-based competition will result in economic stratification of a society based on three consequences of the implementation of ICT. First, there will be displaced employees whose jobs were eliminated and substituted by IT. This is important to note: this is not a scenario of “you can get a raise or you can get a cut”, but, rather, it is a scenario of “sorry, there is nothing for you”. Displaced employees will come from all three organizational levels – operational, tactical, and strategic. Second, there will be a drive to eliminate middle managers – some of them will be eliminated, others will become a part of a strategic level, and yet others will descend to operational level. Appropriate levels of compensation (e.g., increased, decreased, and welfare) will determine the strata of the destination. Finally, there will be a greater disconnect between operational level of a firm, and its strategic level. The reason is simple – the interface, the middle managers, is gone. This stratification of the society based on economics results in strengthening of the cohesiveness of the political, cultural, and religious sub-groups. A general need for an affiliation will drive people towards joining or becoming more active in, what is perceived as, “stable, consistent, and reliable” social group. That, in turn, will lead to “us vs. them” attitude of the members of sub-groups. Interestingly, this attitude will have two directions – internal and external. The internal direction will be pointed against the sub-groups of the homeland economy that are perceived to be responsible for negative changes that took place and impacted the local labor force. The external direction will have its aim on the members of “other” economy(-ies) that are perceived to be responsible for the deterioration of the economic situation in the homeland. This allows us to put forward Negative Inevitable #9, as follows: An introduction of new ICT as a tool of inter-economy competition will result in a compartmentalization of the social environment leading to the increased levels of hostility towards local and external groups perceived to be responsible for the deteriorating economic conditions of negatively impacted subgroups.

Impact of Collaboration Finally, let us consider a case of an inter-economy collaboration – the case of two or more countries working on an ICT project. Unless the economies involved are fairly aligned, from a political, social, and cultural standpoint (e.g., Finland collaborating with Sweden), certain domination-based scenarios are bound to unfold. Culturally, a clash is inevitable and two simple outcomes of such clash are a disintegration of a project, or a submission of a weaker partner to the dominating counterpart. Politically, the impact will be felt by the “junior” member(s) of the project team because the power structure of the weaker member of a team will be under constant pressure of a “stronger” partner. Socially, the impact will be felt if the assignment

Intrinsic Negative Socio-Economic Implications

269

of the roles and responsibilities of a weaker partner does not match the model of a more powerful entity. At this point we can state another Negative Inevitable, #10, as: International collaborations on ICT will face the conflicts associated with and brought by the social, political, and cultural diversity of the collaborating members that will be resolved via the domination of more powerful member of the group.

Investigating Negative Implications of ICT: What Is the Plan? At this point it is time to start contemplating the means by which Negative Inevitables mentioned in this editorial, as well as additional ones yet to be discovered and argued for, could be investigated. We would like to propose a simple, two-step plan of action. First, researchers working in the area relevant to the subject matter must take into consideration, explicitly, negative implications of ICT. And this is not an easy thing to accomplish, for one would have to account for the negatives that take a part of ICT-based glory away. It will, indeed, require a paradigm shift away from the all too familiar “IT is good for development” attitude. Is a sharp knife good? Well, it depends … So, this “it depends” perspective must be embraced. This allows us to formulate Research Suggestion #1: An investigation of positive socio-economic impacts of ICT should also consider, explicitly, collateral negative outcomes. Second, investigators must consider the appropriate targets belonging to the different levels of inquiry. We consider three levels of granularity at which an investigation can be conducted – case studies, frameworks, and grand theories. In our view the three levels relate to each other as follows. Grand theories espouse relationships between general, context-independent constructs, frameworks operationalize the constructs in the context-specific manner, and case studies draw rich pictures of the context from “boots on the ground” perspective. Ideally, there will be a hierarchy and continuity of representation of the results and the data between the levels. Thus, findings of a case study should be possible to frame as a representation of a context-specific construct, and a context-specific construct should be possible to abstract and represent as a more general context-independent construct of a grand theory. If one is to undertake a case study within the context of a single economy, then the local inflections must be identified. Meaning, it is fine and well that the farmers now have the cell phones to get the latest and the most accurate prices for their produce, but what happened to the dealers who used to direct the flow of the goods to the appropriate locations? This allows us to formulate Research Suggestion #2: A case study dedicated to investigating socio-economic impacts of ICT should identify local, specific to the context of the case study, collateral negative outcomes.

270

Quantitative Methodologies using Multi-Methods

In the case of an investigation relying on a framework, a researcher should be able to suggest, first, how a context-specific construct could be generalized to a contextindependent one of a grand theory, and, second, how a context-specific construct may manifest itself in the “boots on the ground” setting. For example, if the results of a case study showed that an introduction of cell phones to a farming community resulted in elimination and substitution of local produce dealers, then what would the appropriate framework-level construct be? We would like to formulate Research Suggestion #3, as follows: A framework-based research should incorporate within its model a contextspecific construct representing collateral negative outcomes of ICT such that: (1) the construct could represent more specific local inflections of a case study-context, and, (2) the construct could be generalized to the level of theory. Finally, if conducting a study backed by a theory, the investigator should attempt to include a general construct representing a collateral negative impact of ICT. Clearly, creating such construct will be driven more by philosophical consideration, then by practice. Nevertheless, the investigator must keep in mind that such a construct should be translatable to the level of a framework, and then to the level of a case study. At this point we can formulate our Research Suggestion #4: A theory-based investigation should incorporate a context-independent construct representing collateral negative outcomes of ICT based on, and resulted from, the substitution and elimination nature of ICT. We provide a brief illustration of the suggestions in Table 27.1.

TABLE 27.1 Research Suggestions Taking into Consideration Negative Impacts of ICT Level Theory

Neoclassical Growth Accounting EconOutput = (Labor, Capital, TFP)

Framework

GDP = (IT_Labor, IT_Capital, TFP)

Case Study

Cell phone usage by farmers in Economy X resulted in increasing level of income due to the improved access to a local pricing system

Modified Neoclassical Growth Accounting EconOutput = (Labor, Capital, TFP, -DisplacedLabor) GDP = (IT_Labor, IT_Capital, TFP, -Displaced_IT_Labor) Cell phone usage by farmers in Economy X resulted in increasing level of income of the farmers due to the improved access to a local pricing system, but it also resulted in elimination and substitution of the services of local produce brokers resulting in ….

Intrinsic Negative Socio-Economic Implications

271

CONCLUSION One of the most important issues that researchers and practitioners must address is that of a chosen granularity of a project. In the case of academics, for example, it concerns the level of investigation and, correspondingly, the level of research questions that are going to be stated. For all intents and purposes this deals with selecting along the “general-to-specific” (e.g., grand theory to case study) spectrum of an inquiry. If, let us say, a research question deals with the finding an impact of X on Y and the impact is found to be significant, then it is only so for the level of the context of the study. If an investigation is context-specific, then the findings also are. What was found positive for the wealthy may not be so for the poor (Harris, 2016). Similarly, practitioners working in the area of ICT deal with projects introducing a change into a particular context, and the context could be very narrow (e.g., level of a department), or it could be very broad (e.g., level of economy). The chosen context carries a corresponding level of scrutiny regarding the success of the project. So, if an implementation of a new software for HR Department was a successful one, then the designation of a “successful project” was based on the stated in advance criteria (e.g., let us say, satisfaction of user requirements) specific to the context of the project. The success of the project at the level of a department may bring negative consequences along the “general-to-specific” spectrum – it may negatively impact an employee of the HR Department, and it could bring some negatives at the level of the firm that houses the department. There is a general similarity between researchers and practitioners working in the area of ICT, and this similarity is in terms of the impact of ICT. Most of the time, the investigators are searching for a positive impact of ICT, and, all the time practitioners have in mind obtaining a positive impact of ICT. The problem is, of course, is that of context. And the question is, at what level? Any introduction of change results in an impact. And it is impossible to have a positive impact without some negative aspects manifesting themselves here and there. If we operate in a single context, which is akin to having a tunnel vision, then we may not see the negatives lurking “here and there”. Interestingly, despite the fact that the tools allowing for a multi-level consideration of the context were available to researchers and practitioners for quite some time (Jackson, 1982; Arnold & Wade, 2015), we don’t see their widespread application.

REFERENCES Arnold, R., & Wade, J. (2015). A definition of systems thinking: a systems approach. Procedia Computer Science, 44, 669–678. https://doi.org/10.1016/j.procs.2015.03.050. Bosamia, M. P. (2013). Positive and negative impacts of information and communication technology in our everyday life. In Proceedings of International Conference on Disciplinary and Interdisciplinary Approaches to Knowledge Creation in Higher Education: Canada & India (GENESIS 2013). Harris, R. (2016). How ICT4D research fails the poor. Information Technology for Development, 22(1), 177–192. https://doi.org/10.1080/02681102.2015.1018115.

272

Quantitative Methodologies using Multi-Methods

Jackson, M. (1982). The nature of “soft” system thinking: the work of Churchman, Ackoff, and Checkland. Journal of Applied Systems Analysis, 9, 17–29. Tarafdar, M., DArcy, J., Turel, O., & Gupta, A. (2015). The dark side of information technology. MIT Sloan Management Review, 56(2), 61–70. Tarafdar, M., Gupta, A., & Turel, O. (2013). The dark side of information technology use. Information Systems Journal, 23(3), 269–275. https://doi.org/10.1111/isj.12015.

Section IV Appendix X THE PURPOSE AND THE SUGGESTED USE OF THE CONTENT IN THIS APPENDIX The content presented in this section of the book is intended to benefit, primarily, researchers working in the area of Information and Communication Technologies for Development (ICT4D). Based on our experience, there are three preliminary questions that a ICT4D investigator must address in the beginning of the paper in order to elicit positive responses of the reviewers and editors, as well as in order to incorporate her work within a nomological network weaved by previous investigations in that area. The first question is:

What is the model/theory of economic development that underlies the inquiry?



An implicit assumption that we relied on in our work is that economic development is a necessary condition for a broader human and societal development to take place – may be an insufficient condition on its own, granted, but a necessary one at that. The content presented in X1 should help an investigator in selecting an appropriate to the subject and context of her study model of economic development.

274

Appendix X

The second question is:

What is the research framework/model of the study?



Fundamentally, research questions of any investigation should not come out of the blue – instead, a set of research questions is generated based on the selected (or constructed) framework (research model) of the study. Needless to say, in the presence of a theory supporting the investigation, the research framework/ model must be demonstrated to be consistent with the selected theory. This way, the investigator is able to show to her readers, reviewers, and editors alike that the research questions of the study were generated based on the research framework/model of the investigation, where the framework is itself consistent with the theory underlying the inquiry. The content presented in X2 may assist an investigator by illustrating the way of selecting or generating a consistent with a theory research framework/model.

X1

Models of Economic Growth

One of the important requirements for selecting a theory of economic development within the context of this text is that it should allow for considering ICT being one of the factors of economic growth. This requirement allows us to remove from consideration such candidates as the early theories of economic development espoused by Adam Smith (where the specialization via division of labor and exchange are the mechanisms for economic growth) and Karl Marx (where public property and planned economy serve as the engines of development). One of the key characteristics of ICT is its ability to offer less developed economies an opportunity to develop rapidly, jumping levels and bypassing the usual milestones (e.g., developing cell networks without prior development of landline networks) – this implies that economies may follow their own unique paths to development. Unfortunately, this makes the Linear Stages of Growth’ economic models (e.g., Modernization Theory) unsuitable for the purpose of supporting our inquiry, because, based on the fundamental assumption of the model, every economy must pass through the same phases (e.g., the traditional society, the preconditions for take-off, the take-off, the drive to maturity and the age of high mass consumption (Rostow, 1962)), stage by stage. It has been established that an implementation of ICT requires development of the labor force – it would appear that we could adopt Structural Change models as an underlying framework for our purposes. After all, steady accumulation of physical and human capital, paired with savings and investments, serve as drivers of economic growth for the proponents of the theory. However, Structural Change models deal with the reallocation of labor from the agricultural to the industrial sector as being the driver of economic growth. This has two important implications. First, implementation of ICT is associated not with industrial, but with post-industrial sector; and second, the policies of shifting labor force from agricultural sector in developing countries have been widely recognized as having an overall negative effect (World Bank, 2000). Thus, we remove this model from further consideration. Implementation of ICT in an economy is often perceived as a collaborative intercountry effort, where, quite often, wealthier counterparts and international agencies help their less developed peers. Perhaps, International Dependence models (Heller, Rueschemeyer, & Snyder, 2009) could be adopted to support our inquiry. However, the main premise of the associated theory is that the less developed countries are taken advantage of by their more developed, wealthier counterparts – consequently, a lack of development is due to the dominance of developed countries over developing countries, and the path to success for developing countries is via the route of ending the economically detrimental relationships with the developed world and by closing their borders and economies to the developed countries. History, however, 275

276

Appendix X1

demonstrated that it is the emphasis on cooperation and trade with the advanced economies (which is contrary to autarkical model espoused by the theory) that resulted in economic success for developing economies (e.g., Hong Kong, Singapore, Taiwan and South Korea, during the 1970s and 1980s) – this resulted in decline of popularity of the theory in 1980s. Unsurprisingly, we consider International Dependence theory as not being suitable for the purposes of our study. Contrary to the perspectives of the International Dependence model, Neoclassical theories (Hahn, 2010) blame domestic, not international issues for the lack of economic development (au contraire, foreign aid and foreign trade are some of the drivers of the development). Instead of external factors, it is internal issues of poor resource allocation, price distortion and corruption that apply the brakes to economic development (Meier & Stiglitz, 2000). We must note that the above mentioned problems, at least in part, are related to the issue of information transparency, which is successfully dealt with via application of ICT. A free market line of thinking within Neoclassical theories is a Traditional Neoclassical Growth model (framework of Neoclassical Growth Accounting, a.k.a. Solow’ model), where economic growth is impacted by quality and quantity of labor, capital investments, and technological progress. However, social, political, and cultural aspects also play an important role (let us recall that this perspective embraces free market), and the economic development could fail to materialize to its fullest extent due to inadequate legal and regulatory framework. Despite its shortcomings, the Neoclassical Growth model fits well with the purposes of our inquiry for the following reasons: • The focus of the model is internal to the economy – the sources of growth reside under the purview of policy makers of a given economy. • The constructs of the model could be uniformly represented across a variety of economies. • The constructs of the model could be objectively reflected by commonly accepted measures. • The model associates the barriers to economic growth with limitations of information channels and information transparency – aspect easily impacted by application of ICT. • The model considers the importance of an adequate governmental services to economic development – and ICT has been shown to be a useful tool in improving quality of governmental services. While Solow’s growth model considers one of the sources of growth – technological progress – to be an exogenous (external) factor, the New Growth theory (Islam, 2004) explicitly links, within the model, technological progress to the production of knowledge as a driver of sustained economic growth. Within this framework, the process of knowledge creation cannot be left to individuals, because the individual would not be able to capture all the gains associated with new knowledge creation. Thus, the role of government and public policies is emphasized as a tool in formation of a human capital and encouragement of foreign private investments in knowledgeintensive industries, such as those associated with ICT. The New Growth theory would seem a perfect candidate to support our inquiry, but two points of concern stop

Appendix X1

277

us from adopting it as a foundation of the study. First, developing economies may not be a good target for a model that emphasizes a knowledge-based growth (Kaur & Singh, 2016), for it has been noted that in less developed economies there are many other factors that are more likely to drive the economic development (Cornwall & Cornwall, 1994; Parash, 2015). Second, any meaningful application of the theory requires representation and capturing of the concept “knowledge creation”, and, given the context of the study, it would be an unsurmountable challenge to overcome. Unlike previously covered models, the theory of Coordination Failure (Hoff, 2000; Kydd & Dorward, 2004) does not have a fixed culprit responsible for the lack of economic development. It is not Adam Smith’s division of labor and market exchange, it is not public property and planned economy of Marx, it is not “must go through” sequence of Linear Stages, it is not a better developed oppressor of International Dependence, it is not reallocation of labor of Structural Change, and it is not absence of capital investments, labor, and technical progress of Neoclassical Growth model, neither is it the knowledge creation of the New Growth theory. Instead, it is a market’s failure to coordinate complementary activities that is responsible for the lack of economic development – simply put, in the absence of coordination, market settles on a sub-optimal result. A possible cure to such situation is the emphasized role of government (because if left on their own, markets settle on local maxima versus converging on a global maximum), and of its supervision over public-driven massive investment program geared toward creating complementarities in the market. From one perspective, Coordination Failure theory is well-suited for the purpose of investigating “anything ICT” – after all, ICT is introduced as a complementarity (e.g., IT complements the skills of an employee) and is intended to produce complementarities along the way (e.g., IT creates and introduces new jobs in newly created sectors – cybersecurity, data analytics, etc.). Yet, from another perspective, this theory is almost impossible to apply for our purposes, for the following reasons: • The scope of ICT-related complementarities within an economy is hard to define, to uniformly describe, and impossible to measure on a common scale. • The government-centric “central planning” approach to managing complementarities is hardly applicable to the context of the study. • The “big push” model of government-supervised investments may not be supported but what is actually happening in developing economies. For one, we do not, and did not, see a large scale “big push” investments in ICT. There is no single agreed-upon theory of development – each model considers a few relevant dimensions without claiming exclusivity of knowledge or a gnostic insight on the topic. Behind the diversity of perspective on the issue of economic growth there are three common questions that every model tries to deal with. The first question is how to explain economic development over time. The second question is how to identify barriers to growth. And the third question is what is the appropriate role of a government in the process. In order to support our line of inquiry, we identified the framework of Neoclassical Growth as a suitable candidate – we hope that, given the reviewed above alternatives,

278

Appendix X1

our readers would concur with the selection made. In simple terms, the framework of Neoclassical Growth Accounting explains the economic growth in terms of capital, labor, and technological progress, where governmental policies of liberalization, stabilization and privatization are central elements of the development effort. We must note that the selected theoretical support is not exclusive to other perspectives on development. For example, the Neoclassical Growth model is perfectly compatible with and complementary to the Development as Freedom’ perspective of Amartya Sen (1999). Let us consider two aspects of the selected model – investments in ICT as a factor of economic development and macroeconomic growth. ICT Capabilities created as a result of investments in ICT are perfectly available, via creating new and optimizing existing information channels, of addressing some of the issues associated with political freedoms and transparency of the relationships between people and people, people and government, and people and social and political groups. Macroeconomic growth, on the other hand, is capable of addressing a freedom of opportunity via generating additional wealth and affording a greater access for poor to financing and credit. Similarly, additional wealth generated as a result of economic development may offer a greater level of economic protection for the poor via more money being available to allocate to income supplements and unemployment benefits. It is perfectly reasonable, consequently, to see Neoclassical Growth model as describing how to generate wealth, and to see Sen’s model as describing how to allocate generated wealth in order to address important problems of social and human development.

REFERENCES Cornwall, J. and Cornwall, W. (1994). Growth Theory and Economic Structure. Economica, 61(242), 237–251. Hahn, F.H. (2010). Neoclassical Growth Theory. In: Durlauf S.N., Blume L.E. (eds.) Economic Growth. The New Palgrave Economics Collection. Palgrave Macmillan, London. Heller, P., Rueschemeyer, D. and Snyder, R. (2009). Dependency and Development in a Globalized World: Looking Back and Forward. Studies in Comparative International Development (SCID), 44(4), 287–295. Hoff, K. (2000). Beyond Rosenstein-Rodan: The Modern Theory of Underdevelopment Traps. World Bank Development Economics Conference 2000. Islam, N. (2004). New Growth Theories: What Is in There for Developing Countries? The Journal of Developing Areas, 38(1), 171–212. Kaur, M. and Singh, L. (2016). Knowledge in the Economic Growth of Developing Economies. African Journal of Science, Technology, Innovation and Development, 8(2), 205–212. Kydd, J. and Dorward, A. (2004). Implications of Market and Coordination Failures for Rural Development in Least Developed Countries. Journal of International Development, 16, 951–970. Meier, G. and Stiglitz, J. (2000). Frontiers of Development Economics: The Future in Perspective. New York: World Bank and Oxford University Press. Parash, U. (2015). Factors Affecting Economic Growth in Developing Countries. Major Themes in Economics, 17, 37–54. Rostow, W. W. (1962). The Process of Economic Growth. Clarendon Press, Oxford. Sen, A. (1999). Development as Freedom (1st ed.). Oxford University Press, New York, NY. World Bank. (2000). Entering the 21st Century – World development report 1999/2000. Oxford University Press, New York, NY.

X2

A Model of the SocioEconomic Impact of ICT

An overview provided in the previous part of the text, capped by the selection of the preferred theoretical framework, allows us to formulate a conceptual model according to which ICT contributes to economic development – it is as follows: (ICT Labor, ICT Capital/Investments, ICT-driven Technological Progress) → Macroeconomic Growth. A reliance on such grand theory as Neoclassical Growth model allows for linking the factors of economic growth to the growth itself, but it does not shed any light on the mechanisms by which, let us say, labor and investments work their ways to contributing to macroeconomic growth. As preamble to the discussion of the link between ICT and macroeconomic growth we need to establish a connection between macroeconomic development and factors of growth – investments and labor. The obvious connection would be that a higher level of economic development allows for higher levels of investments and higher quality of the labor force – this, over time, results in the accumulation of labor – as well as a build-up of capital-related ICT infrastructure. But, on their own, accumulated capital, investments, and available labor are not very useful – they are simply resources waiting to be taken advantage of. Simply put, increasing levels of investments in ICT fueled by economic development, applied on top of the existing socio-technical ICT infrastructure, create increasing levels of ICT Capabilities. The Networked Readiness Index (NRI) serves as an established measure of the ICT Capabilities, or, drivers, of socio-economic development (World Economic Forum, 2016). Since its development in 2002 the NRI framework has been highly regarded as an authoritative source of data and an assessment tool geared towards global leaders, practitioners, policy makers (Kirkman, Osorio, & Sachs, 2002), which was also available to academic researchers (Indjikian & Siegel, 2005; Ifinedo, 2005; Vehovar et al., 2006; Wielicki & Arendt, 2010; Chanyagorn & Kungwannarongkun, 2011; Milenkovic et al., 2016; Samoilenko & Osei-Bryson, 2017, 2019). The NRI framework used in this text is reflected by a single composite index comprised of four sub-indexes – three drivers (Environment, Readiness, and Usage sub-indexes) and one Impact sub-index. The sub-indexes are represented by 10 subcategories. The values of the index, sub-indexes, and sub-categories are represented by scores ranging from 1, being the lowest, to 7, being the highest – this allows not only for a cross-country comparison, but also for comparison of each country vis-àvis itself over time. 279

280

Appendix X2

One of the underlying principles of the framework is that “NRI should provide clear policy guidance” (World Economic Forum, 2016, p. xi) to its users, which is not an easy undertaking to implement because within the framework “…the complex relationships between ICTs and socio-economic performance are not fully understood and their causality not fully established” (Di Battista, Dutta, Geiger, & Lanvin, 2015, p.4). Unsurprisingly, the framework of NRI conceptualizes and captures ICT Capabilities in terms of three Drivers – Environment, Readiness, and Usage. It is worth noting again – ICT capabilities are created on top of the existing socio-technical infrastructure. This allows us, so far, to establish and justify the following chain of links: Economic Development → Investments in ICT (applied to existing sociotechnical ICT infrastructure) → ICT Capabilities → Economic Development. We must note that while it is inevitable that investments in ICT would work via creation of ICT Capabilities, investigators have options in representing the ICT Capabilities construct (Zuppo, 2012; Kleine, 2013). In this chapter, we selected the framework of NRI for the purpose because it offers an established representation that is widely accepted by scholars and practitioners, and also could be applied to our context – two different sets of economies. However, in a narrower context investigators have been known to “unpack” the meaning and representation of the construct to elucidate transformation paths by which investments make their ways to becoming micro- and macroeconomic benefits. For example, Samoilenko (2016) traced the investments in ICT to economic development path via a chain of links starting from the introduction of Telecom products, to decreasing cost of acquisition and utilization of Telecom products, and then to the resultant increase in the level of consumption. Another advantage afforded by a reliance on the NRI framework is that it includes an Impact sub-index, which represents and measures social and economic impacts of the Drivers – ICT Capabilities created as a result of investments in ICT. However, within the framework “…the complex relationships between ICTs and socio-economic performance are not fully understood and their causality not fully established” (Di Battista, Dutta, Geiger, & Lanvin, 2015, p.4). Consequently, Drivers of NRI are not causally linked to Impact – this poses a question: How do Drivers produce Impact? An answer was offered by Samoilenko (2018), who suggested that ICT Capabilities (e.g., Drivers) must enable some sort of a socio-economic engine that actually produces Impact, and the consequent framework was developed by Samoilenko and Osei-Bryson (2019). Importantly, this model of “Drivers → Engine → Impact” was then reconciled with the Neoclassical Growth model using a framework of sustainable macroeconomic impact of investments in ICT (Samoilenko, 2016) – the resultant model is depicted by Figure X2.1. The model depicted in Figure X2.1 supports the development cycle of the study, which makes the theoretical foundation of our investigations to be consistent not

Appendix X2

281

FIGURE X2.1  Framework of NRI and Neoclassical Growth model (Samoilenko & OseiBryson, 2021).

only with the established NRI framework, but with the grand theory of Neoclassical Growth as well, without making it inconsistent with other perspectives on development (e.g., Sen, 1999). In 2020 the NRI released a new model that, while maintaining a continuity with the previous version, is better geared towards reflecting current and future ICTrelated issues. The framework still consists of three drivers and one impact, but the names and representations of sub-indexes change – the new model is comprised of four pillars: Technology, People, Governance, and Impact. The changes in names or representations do not impact the overall structure of the Drivers → Engine → Impact model of Samoilenko and Osei-Bryson (2019) – we present their framework with the updated version of NRI in Figure X2.2.

FIGURE X2.2  Drivers → Engine → Impact model using the 2020’ version of NRI.

282

Appendix X2

REFERENCES Chanyagorn, P. and Kungwannarongkun, B. (2011). ICT Readiness Assessment Model for Public and Private Organizations in Developing Country. International Journal of Information and Education Technology, 1(2), 99–106. Di Battista, A., Dutta, S., Geiger, T. and Lanvin, B. (2015). The Networked Readiness Index 2015: Taking the Pulse of the ICT Revolution, Global Information Technology Report 2015, pp. 3–28. Ifinedo, P. (2005). Measuring Africa’s E-readiness in the Global Networked Economy: A Nine-Country Data Analysis. International Journal of Education and Development using ICT, 1(1), 53–71. Indjikian, R. and Siegel, D. (2005). The Impact of Investment in IT on Economic Performance: Implications for Developing Countries. World Development, 33(5), 681–700, Kirkman, G. S., Osorio, C. A. and Sachs, J. D. (2002). The Networked Readiness Index: Measuring the Preparedness of Nations for the Networked World. The Global Information Technology Report: Readiness for the Networked World, 10-30. Oxford University Press, New York, NY. Kleine, D. (2013). Technologies of Choice? ICTs, Development, and the Capabilities Approach. The MIT Press, Cambridge, MA, London, England. Milenkovic, M. J., Brajovic, B., Milenkovic, D., Vukmirovic, D. and Jeremic, V. (2016). Beyond the Equal-weight Framework of the Networked Readiness Index: A Multilevel I-Distance Methodology. Information Development, 32(4), 1120–1136. Samoilenko, S. (2016). Where Do Investments in Telecoms Come from? Developing and Testing a Framework of Sustained Economic Impact of Investments in ICT. Journal of Information Technology for Development, 22(4), 584-605. Samoilenko, S. (2018). Socio-Economic Impact of ICT-Enabled Public Value in Sub-Saharan Economies. In Proceedings of 6th Mediterranean Interdisciplinary Forum on Social Sciences and Humanities, MIFS 2018, 24–25 May 2018, Barcelona, Spain. Samoilenko, S. and Osei-Bryson, K.M. (2017). An Analytical Framework for Exploring Context-Specific Micro-Economic Impacts of ICT Capabilities. Journal of Information Technology for Development, 24(2), 633–657. Samoilenko, S. and Osei-Bryson, K.M. (2019). Representation Matters: An Exploration of the Socio-Economic Impacts of ICT-Enabled Public Value in the Context of Sub-Saharan Economies. International Journal of Information Management, 49, 69–85. Samoilenko, S. and Osei-Bryson, K.M. (2021). ICT Capabilities and the Cost of Starting Businesses in Sub-Saharan African Economies: A Data Analytic Exploration, Journal of Global Information Technology Management, 24(1), 7–36. Sen, A. (1999). Development as Freedom (1st ed.). Oxford University Press, New York, NY. Vehovar, V., Sicherl, P., Hüsing, T. and Dolnicar, V. (2006). Methodological Challenges of Digital Divide Measurements. The Information Society, 22(5), 279–290. Wielicki, T. and Arendt, L. (2010). A Knowledge-Driven Shift in Perception of ICT Implementation Barriers: Comparative Study of US and European SMEs. Journal of Information Science, 36(2), 162–174. World Economic Forum. (2016). Networked Readiness Dataset. Available on-line at: http://reports. weforum.org/global-information-technology-report-2016/networked-readiness-index/ Zuppo, C.M. (2012). Defining ICT in a Boundaryless World: The Development of a Working Hierarchy. International Journal of Managing Information Technology, 4, 13–22.

Index Italicized pages refer to figures and bold pages refer to tables.

A Agglomerative hierarchical clustering, 10 Annual telecom investments, 229 A priori, 14 Arcelus, F. J., 117, 119 ARM, see Association rules mining Arocena, P., 117, 119 Association rules mining (ARM), 14–16 data set for, 15–16 for discovering common causal structures, 177, 177–178 empirical investigation of ICT capabilities association rule mining, 188 data analysis, results of, 192–194, 193–194 interpretation of results of data analysis, 197–198 heterogeneous sample (groupings are given), 34–35, 43–46 DEA models inputs, generating, 45 DEA models outputs, generating, 45 expanded data set, 43 itemsets in, 34, 44 level of averaged relative efficiency, generating, 45 level of top-split variables, generating, 45 received categorization, generating, 45–46 sub-sets of sample, 34–35 top-split variables of DTI, 34 heterogeneous sample (groupings are not known), 56–57 contextual variables and variables of DEA model, itemsets based on, 75 contextual variables or variables of DEA model, n-clusters in terms of, 76 itemsets in, 65–66 multiple M analyses, 66 using intrinsic and contextual variables, 57 using only contextual variables, 56–57 using only intrinsic variables, 56 variables of DEA model, itemsets based on, 75 variables of DEA model, n-clusters and, 76 homogeneous sample, 31–32 data conversion, 31–32

efficient vs. inefficient DMU, 32 purpose of, 32 Market Basket Analysis, 14–16; see also Market Basket Analysis vs. neural networks, 14 socio-economic impacts of ICT-enabled public value in SSA, 209, 224, 224–225 for socio-economic outcomes of ICT capabilities in SSA, 148, 152, 152–153 Athey, S., 227 Ayabakan, S., 157

B Banker, R. D., 171, 179 Barbosa, A., 171 Bardhan, I. R., 157 Benchmarking, 157–171 of business processes, 159 common causal structures, non-obvious, 161–162 concept of, 158 data envelopment analysis for, 157 via Enterprise Systems, 161 problem, conceptualization of, 158–159 Black box model association rules mining, 14 neural networks, 12, 13, 14 Black box transformative capacity model (BBTM), 113, 114 BPR, see Business process re-engineering Brusca, I., 171 Bugamelli, M., 227 Business process re-engineering (BPR), 161

C CA, see Cluster analysis Campos, N., 235 Casual model, DTI, 38 Categorization, 3 CBSP, see Cost of business start-up procedures CDTI, see Classification decision trees induction Ceteris paribus, 157, 185, 201 Change in efficiency (EC), 98 Change in technology (TC), 98 Charnes, A., 245

283

284 Classification decision trees induction (CDTI), 10–12 categorization, 11 reasons to use, 10–11 target variable, 11 top-level nodes, 12 top-level split, 11 Cluster analysis (CA), 9–10 agglomerative hierarchical clustering, 10 divisive hierarchical clustering, 10 empirical investigation of ICT capabilities, 187 data analysis, results of, 189, 190 interpretation of results of data analysis, 194 heterogeneous sample (groupings are not known), 51–52, 55–56, 69–71, 77–78 actionable variables, 55 based on DEA model, 69–70 based on DEA model and contextual variables, 70–71, 82–83 contextual factors, 73–74 endemic factors, 74 intrinsic variables, 55 k-means clustering, 60, 64, 78 options for, 55–56 purpose of applying, 77 time-series data in, 52 heterogeneous subgroups, 9 information technology investment utilization in TE, 229–230, 232–234, 233, 234 K-means, 10 socio-economic impacts of ICT-enabled public value in SSA, 209, 210, 210–211, 211 unguided technique, 9 Cobb–Douglas production function, 224, 226 Common causal structures, discovering, 157–179 benchmarking, process of; see also Benchmarking enhancing, 161–162 hypothetico-deductive logic, 164, 164–165 problem conceptualization of, 158–159 contextual translation of meta-causal structures into specific structures, 162 illustrative example (application to SubSaharan economies), 172–178 association rule mining (phase 5), 177, 177–178 data envelopment analysis (phase 3), 173–176, 174, 175 decision tree induction (phase 4), 176, 176 Malmquist index, DEA to calculate values of, 174–175 NRI framework (phase 1), 172, 172–173

Index partition set of decision making units (phase 2), 173 non-obvious, 157, 161, 162, 163, 164, 166 overview, 157–158 proposed methodology, 168–171 association rules mining, 170–171 data envelopment analysis, 170 decision tree induction, 170 hybrid DEA/DTI methodology, 171 justification and benefits of, 169–171 novelty and value of, claims for, 179 2-stage DEA+OLS model, 171 research problem and research questions of study, 159–168 benchmarking, impact of, 161 context-specific factors, 160–161 input–output conversion, 161 process management and improvement techniques, insufficient, 161 pure input–output process, 160 scope of, 166 Complementarity of investments, 225–227 Confidence, 15 Constant return to scale (CRS), 98, 99, 117, 231 Farrel Input-Saving Measure of Efficiency, 243 relative efficiencies of DMU, 252 Contextual independent variable, 39 Coordination Failure theory, 277 Coricelli, F., 235 Cornia, G. A., 235 Cost of business start-up procedures in SSA, ICT capabilities impact on, 185–201 association rule mining, 188 interpretation of results, 198 results of, 192–194, 193–194 cluster analysis, 187 results of, 189, 190 data envelopment analysis, 188, 191 results of, 191, 192 decision tree induction, 188 interpretation of results, 196 results of, 189–190, 190 NRI framework, 186 ordinary least squares regression, 188 interpretation of results, 197 results of, 191–192, 192 socio-economic impacts, 185–186 CRS, see Constant return to scale

D Data envelopment analysis (DEA), 16–18, 73, 109–111 and base-oriented model, 17 for benchmarking, 157

Index for discovering common causal structures, 173–176, 174, 175 discriminatory power of, 245–257; see also Discriminatory power of DEA for disparities in socio-economic outcomes of ICT capabilities in SSA, 147, 147–148, 150–152, 151 empirical investigation of ICT capabilities, 188, 191 data analysis, results of, 191, 192 interpretation of results of data analysis, 196 functional similarity of DMU, 245–246 fundamental assumptions of, 109–110, 245 heterogeneous sample (groupings are given), 43, 48–49 contextual variables, 47 inputs-outputs of, sub-groups associated with, 48 orientation of model, 48 simulated data in, 50 heterogeneous sample (groupings are not known), 69–71, 74–75, 78–79 CA based on, 69–71 level of performance of DMU, 83–84 logical partitioning, 71, 78, 83–84 low-level cluster, data sets for, 85–86 physical partitioning, 71 information technology investment utilization in TE, 228–229, 231–232, 232 input-oriented model, 17 inputs and outputs of, 16–17, 19 Malmquist Index, 17–18 vs. multivariate regression, 18–20 as non-parametric analytic tool, 109, 114, 119 non-parametric method, 245 non-parametric method, 245 output-oriented model, 17 point-in-time method, 17 propose of, 246 for relative efficiency of decision-making units, 16–18 socio-economic impacts of ICT-enabled public value in SSA, 209, 212, 213 variable return to scale, 147 DEA, see Data envelopment analysis Decision-making units (DMU) efficient vs. inefficient, handling of, 32 homogeneity assumption, 110 relative efficiency of, 16–18, 109–111; see also Data envelopment analysis scale heterogeneity, 110–111 transformative capacity – capability of, 47, 111 Decision support system (DSS), 90–93 architecture or design of, 93, 93, 103–104

285 decision rules generating by decision tree analysis, 100 externally oriented functionality, 90–91 illustrative application of, 94–103 competitive business environment, nature of, 95, 95–96 complementarities between relevant variables, 101–102 DEA model, 97–98 differences in relative efficiency, factors associated with, 99, 99–101 heterogeneity of business environment, factors for, 96, 96–97 input–output process, increasing efficiency of, 102, 102–103 relative efficiency, differences in, 97, 97–101, 98 SAS’ Enterprise Miner for, 94–95 variables used in, 95 internally oriented functionality, 92–93 sequential method utilization within design of, 93 Decision tree analysis decision rules generated by, 100 information technology investment utilization in TE, 230–231, 234–235 Decision trees induction (DTI) classification see Classification decision trees induction for discovering common causal structures, 176, 176 for disparities in socio-economic outcomes of ICT capabilities in SSA, 148, 152, 152–153 empirical investigation of ICT capabilities, 188 data analysis, results of, 189–190, 190 interpretation of results of data analysis, 195 heterogeneous sample (groupings are given), 33–34, 48 casual model, 38 data set contains contextual variables, 42–43 data set is comprised of variables of DEA model, 42 dominance threshold, 33–34 pure nodes, 33 sub-sets in sample, role of, 34 target variable, 33 top-level splits, 48 heterogeneous sample (groupings are not known), 52, 78 CA-based target variable, 65 data set comprises variables of MR model and contextual variables, 60–61

286 data set limited to variables of MR model, 60 membership of node, 83 multi-dimensional analysis, 52 a priori target variable, 65 vs. neural networks, 12 socio-economic impacts of ICT-enabled public value in SSA, 209, 211, 211 Developed countries, 221 Developing countries, 221 Discriminatory power of DEA, in presence of sample heterogeneity, 245–257 cluster membership, 251 DEA models clusters based on input-oriented model, comparison of, 252 clusters based on output-oriented model, comparison of, 253 list of variables for, 248 output of clustering, 250 overview, 245–247 pairs of rules describe efficiency categories, 257 proposed methodology to study, 247–255 data set of illustrative example, 247–248 DEA models, list of variables for, 248 description of methodology, 248–155 parameters in, steps required for initialization of, 248 relative efficiency categories, 253–254, 254–255 relative efficiency status of DMU, determining, 251–253 structural homogeneity status of data set, determining, 249–251 Divisive hierarchical clustering, 10 DMU, see Decision-making units Dollery, B. E., 171 Dominance threshold, 33 DSS, see Decision support system DTI, see Decision trees induction Dyson, R.G., 111

E Economic development, model or theory of, 273 Economic growth, model of, 275–278 Coordination Failure theory, 277 implementation of ICT, 275 International Dependence models, 275–276 Linear Stages of, 275 Neoclassical Growth model, 276, 278 New Growth theory, 276–277 Solow’s growth model, 276 Edgeworth, F. Y., 225 Emerging economies, 222

Index Empirical investigation of ICT capabilities and CBSP, in SSA, 183–201 cost of business start-up procedures, 185–201; see also Cost of business start-up procedures, in SSA data acquisition, 188–189 data analysis, results of, 189–194 association rule mining, results of, 192–194, 193–194 cluster analysis, 189, 190 data envelopment analysis, results of, 191, 192 decision tree induction, 189–190, 190 ordinary least squares regression, results of, 191–192, 192 interpretation of results, of data analysis, 194–198 association rule mining, 197–198 cluster analysis, 194 data envelopment analysis, 196 decision tree induction, 195 ordinary least squares, 197 proposed methodology, 187–188 association rule mining, 188 cluster analysis, 187 data envelopment analysis, 188, 191 decision tree induction, 188 ordinary least squares regression, 188 research framework and research questions, 186, 186–187 results of study, 198–200 small and medium enterprises, 184–185 socio-economic impacts, 186 framework for, 184 Externally oriented functionality, 89

F Farrel Input-Saving Measure of Efficiency, 231 constant return to scale, 243 non-increasing returns to scale, 244 variable returns to scale, 243 Feedback-type mechanism, 89 function(s) Cobb–Douglas production, 224, 226 input–output transformation, 102 neoclassical production, 129, 224 production, 226 transcendental logarithmic production, 226 transfer, 114

G Gera, S., 227 GITR, see Global Information Technology Report Giuri, P., 227

Index Global Information Technology Report (GITR), 149 Growth accounting, 224–225

H Heterogeneous sample (groupings are given), methodological modules, 21, 33–50, 81–86 association rules mining, 34–35, 43–46 DEA models inputs, generating, 45 DEA models outputs, generating, 45 expanded data set, 43 itemsets in, 34, 44 level of averaged relative efficiency, generating, 45 level of top-split variables, generating, 45 received categorization, generating, 45–46 sub-sets of sample, 34–35 top-split variables of DTI, 34 data envelopment analysis, 43, 48–49 contextual variables, 47 inputs-outputs of, sub-groups associated with, 48 orientation of model, 48 simulated data in, 50 decision trees induction, 33–34, 48 casual model, 38 data set contains contextual variables, 42–43 data set is comprised of variables of DEA model, 42 dominance threshold, 33–34 pure nodes, 33 sub-sets in sample, role of, 34 target variable, 33 top-level splits, 48 multivariate regression creating new, using contextual independent variables, 39 using causal model, 38–39 neural networks, model of transformative capacity, 49–50 purpose of, 33, 37, 41–42, 47 sources of heterogeneity of sub-sets, identifying, 37 Heterogeneous sample (groupings are not known), methodological modules, 21, 51–86 applicability of, 51 association rules mining, 56–57 contextual variables and variables of DEA model, itemsets based on, 75 contextual variables or variables of DEA model, n-clusters in terms of, 76 itemsets in, 65–66

287 multiple M analyses, 66 using intrinsic and contextual variables, 57 using only contextual variables, 56–57 using only intrinsic variables, 56 variables of DEA model, itemsets based on, 75 variables of DEA model, n-clusters and, 76 cluster analysis, 51–52, 55–56, 69–71, 77–78 actionable variables, 55 based on DEA model, 69–70 based on DEA model and contextual variables, 70–71, 82–83 contextual factors, 73–74 endemic factors, 74 intrinsic variables, 55 k-means clustering, 60, 64, 78 options for, 55–56 purpose of applying, 77 time-series data in, 52 data envelopment analysis, 69–71, 74–75, 78–79 CA based on, 69–71 level of performance of DMU, 83–84 logical partitioning, 71, 78, 83–84 low-level cluster, data sets for, 85–86 physical partitioning, 71 decision trees induction, 52, 78 CA-based target variable, 65 data set comprises variables of MR model and contextual variables, 60–61 data set limited to variables of MR model, 60 membership of node, 83 multi-dimensional analysis, 52 a priori target variable, 65 illustrative scenario for, 63 multivariate regression, 59–62 data sets, types of, 59 interpretation of interaction term in, 61–62 top-split variables, 61 neural networks, 84–85 high-level cluster, creating model of, 84 high-level cluster using model of low-level cluster, simulation of outputs of, 85 low-level cluster, creating model of, 84 low-level cluster using model of high-level cluster, simulation of outputs of, 84–85 purpose of, 51, 55, 59, 64, 77 Homogeneous sample, methodological modules, 21 association rules mining, 31–32 data conversion, 31–32 efficient vs. inefficient DMU, 32 purpose of, 32

288 data envelopment analysis, 31 purpose of, 31 Hoskisson, R., 222, 234 Human Development Index, 221 Hybrid DEA/DM-based decision support system, see Decision support system Hypothetico-deductive logic (HDL), 164, 164–165

I ICT, see Information and Communication Technologies ICT capabilities, 125–128 business readiness, 139 data envelopment analysis, application of application of, 131–132 benchmark group for efficiency, 136, 136 input and output components of, 131 Malmquist Index score, 135, 135 relative efficiency score, 135 results of, 134–137 decision tree-based analysis, 133–134 results from, 137–139 description of data, 134 differences between groups of economies in terms of, 138 empirical Investigation of, 183–201; see also Empirical investigation of ICT capabilities environment-related, 139 impact of, 125–128 context-specific, 127 on microeconomic outcomes, 126–127 sustainable, 125–126, 126 neoclassical growth accounting model, 128–129, 129 Networked Readiness Index, 127–128 socio-economic outcomes in SSA, 143–155; see also ICT capabilities in SSA for sustainable impact, 125–126, 126 theoretical framework and research model, 125–126, 126, 128–131 ICT capabilities in SSA, disparities in socioeconomic outcomes of, 143–155 data acquisition, 149–150 data analysis, results of, 150–154 Networked Readiness Index, 145, 145 overview, 143–145 proposed methodology to study, 146–148 association rule mining (phase 3), 148, 152, 152–153 benefits and justifications for, 148 data envelopment analysis (phase 1), 147, 147–148, 150–152, 151 decision tree induction (phase 2), 148, 152, 152–153

Index research framework for, 145, 145 research questions and null hypotheses of study, 148–149 ICT4D, see Information and Communication Technologies for Development ICT-enabled public value in SSA, socioeconomic impacts of, 203–218 data acquisition, 209–210, 210 data analysis, results of, 210–217 methodology to investigate, 208–210 association rule mining, 209, 224, 224–225 cluster analysis, 209, 210, 210–211, 211 data envelopment analysis, 209, 212, 213 decision tree induction, 209, 211, 211 ordinary least squares regression, 209, 212, 213 overview, 203–204 research framework to study, 204–207 ICT capabilities impact, 205–206 NRI framework, 204–205, 205 research questions on study of, 207–208 results of study on, 216–217 ICTF, see Information and Communication Technologies Task Force IMF, 246 Information and Communication Technologies (ICT) advanced-for-the-context, 263 artificial intelligence-enabled, 261 capabilities, 125–128; see also ICT capabilities competing with others, implications of, 267–268 contribution to macroeconomic growth, 222 elimination and substitution conditions for, 262–263 routes of, 261–262 impact of, 125 general model of, 126, 126–127 negative, research suggestions for, 270 sustainable, 125–126 inter-economy collaboration, impact of, 268–269 intrinsic negative socioeconomic implications of, 259–271 investments in, 115 macroeconomic impact of, 129, 130 and state of ICT represented by NRI, 130, 130 microeconomic outcomes, 127–128, 140 component variables of, 134 DEA outputs, 131 income-related, 139 negative implications of, investigating, 269–270 platform, message, and target, 264–267

289

Index pragmatics and ethics of implementation of, 263 role of, 125 socio-economic impacts of dimensions of, 264 overview, 259–260 tools, machines, and, 260–261 Information and Communication Technologies for Development (ICT4D), 127, 128, 141, 273 Information and Communication Technologies Task Force (ICTF), 120 Information technology investment utilization in TE, contributing factors to, 221–240 contribution of study on, 238–239 data acquisition, 227 data analysis, results of, 231–235 efficiency of, general factors contributing to, 237 investment in telecoms, 223 methodology to study, 228–235 cluster analysis, 229–230, 232–234, 233, 234 data envelopment analysis, 228–229, 231–232, 232 decision tree analysis, 230–231, 234–235 overview, 221–223 theoretical framework to study, 224–227 growth accounting, 224–225 theory of complementarity, 225–227 transition economies, definition of, 221 Input-output conversion process, 157, 168 black box model of, 22 outcomes of, 161 Internally oriented functionality, 89 International Dependence models, 275–276 International Telecommunication Union (ITU), 115, 227, 247 Itemset, 14 ITU, see International Telecommunication Union

J Just-in-time (JIT) manufacturing system, 262

K Kaufmann, D., 207 K-means, 10 k-means clustering, 60 Kraay, A., 207

L Lambertini, L., 227 Least-developed countries, 221 Lift, 15 Lima, S. C. D., 171

Lin, P., 227 Logical partitioning, 71 Loukis, E., 227

M Malmquist Index (MI), 17–18, 135, 147, 174–175 change in efficiency, 18, 98 change in technology, 18, 98 sources of change, identification of, 17 technology change component, 18 Market Basket Analysis (MBA), 14–15 confidence, 15 itemset, 14 lift, 15 A priori, 14 purpose of, 14 support, 15 Mastruzzi, M., 207 Methodological modules, framework for, 21–26 development of with consideration of input-output model, 26 without consideration of input-output model, 25 insights and limitations of developed methodological modules, 23–24 of single methods, 22 of two-method combinations, 22–23 Model(s); see also specific models black box, 12, 13, 14 black box transformative capacity model, 113, 114 casual, 38 economic growth, 275–278 International Dependence, 275–276 Neoclassical Growth, 276 neoclassical growth accounting, 128–129, 129 Product Life Cycle, 112 Multivariate regression (MR), 18–20 benefits of, 20 vs. data envelopment analysis, 18–20 heterogeneous sample (groupings are given) creating new, using contextual independent variables, 39 using causal model, 38–39 heterogeneous sample (groupings are not known), 59–62 data sets, types of, 59 interpretation of interaction term in, 61–62 top-split variables, 61 inputs and outputs of, 19 interpretation in, 20 linear relationship, 19 partial correlation, 19 purpose of, 18

290 N Natarajan R., 171, 179 Neoclassical Growth Accounting, 201, 229, 270, 276, 278 framework of, 140, 183 integrating NRI framework with, 184 simplicity of, 225 framework of Samoilenko consistent with, 129 macroeconomic impact of investments in ICT based of, 129, 129 modified, 270 Neoclassical Growth model, 276 Networked Readiness Index (NRI), 127–128, 130, 141, 145, 145 cost of business start-up procedures in SSA, ICT capabilities impact on, 186 for discovering common causal structures, 176, 176 socio-economic impacts of ICT-enabled public value in SSA, 204–205, 205 Neural networks (NN), 12–13 vs. association rules mining, 12, 13 black box model, 12, 13, 14 vs. decision trees induction, 12 heterogeneous sample (groupings are given) model of transformative capacity, 49–50 heterogeneous sample (groupings are not known), 84–85 high-level cluster, creating model of, 84 high-level cluster using model of low-level cluster, simulation of outputs of, 85 low-level cluster, creating model of, 84 low-level cluster using model of high-level cluster, simulation of outputs of, 84–85 hidden nodes, 12–13 input nodes, 12, 13 output nodes, 12, 13 process of training of, 13 New Growth theory, 276–277 New Public Management (NPM), 103 NIRS, see Non-increasing returns to scale NN, see Neural networks Non-increasing returns to scale (NIRS), 98, 99, 117, 231 Farrel Input-Saving Measure of Efficiency, 244 relative efficiencies of DMU, 252 NPM, see New Public Management NRI, see Networked Readiness Index

O OLSR, see Ordinary least squares regression Operational excellence, 160, 163; see also Process improvement

Index benchmarking and, 161–162 direct or indirect inputs for, 157 execution of business processes, 159 input–output conversion, 161 Ordinary least squares regression (OLSR) empirical investigation of ICT capabilities, 188 data analysis, results of, 191–192, 192 interpretation of results of data analysis, 197 socio-economic impacts of ICT-enabled public value in SSA, 209, 212, 213 Osei-Bryson, K. M., 113, 116, 187

P Pagano, P., 227 Partial correlation, 19 Physical partitioning, 71 Piatkowski, M., 233 Platform, definition of, 264 Point-in-time method, 17 Popov, V., 235 Process improvement; see also Operational excellence deductive approaches, 166 hypothetico-deductive approaches to, 166 initiatives, 179 limited importance to, 167 techniques, insufficient, 161 Productivity-driven organization, 89 Product Life Cycle (PLC) model, 112 Public value, 203 ICT-enabled public, impact of, 203–204, 217; see also ICT-enabled public value in SSA influence on socio-economic development of SSA, 204 representation of, 207

R regression multivariate see Multivariate regression ordinary least squares, 188 192, 191–192, 197 Tobit, 114 Relative inefficiency in heterogeneous samples, determining Sources of, 109–120 of decision making units, 109–111 illustrative data set, 115–119 application of methodology on, 116–119 DEA model, variables selected for, 116 description of, 115–116 NN simulation, 118, 118 outputs for each cluster based on BBTM, 118–119

291

Index relative efficiency status of DMU, determine, 117–118, 118 scale heterogeneity status of data set, evaluation of, 116–117 proposed methodology for problems associated with DEA, 111–155, 112 black box model of transformative capacity of each cluster, generating, 113 DEA scores, modeling of, 113–114 motivation for methodology, 113–115 nonhomogenous DMU, problem is associated with, 111 return-to-scale assumptions of DEA, problem is associated with, 111–112 simulated sets of outputs for each cluster, generating, 113 sources of relative inefficiency of DMU, determining, 113 Research framework or model of study, 274 Rosenkranz, S., 227

Socio-economic outcomes of ICT capabilities in SSA, disparities in, 143–155 data acquisition, 149–150 data analysis, results of, 150–154 Networked Readiness Index, 145, 145 overview, 143–145 proposed methodology to study, 146–148 association rule mining (phase 3), 148, 152, 152–153 benefits and justifications for, 148 data envelopment analysis (phase 1), 147, 147–148, 150–152, 151 decision tree induction (phase 2), 148, 152, 152–153 research framework for, 145, 145 research questions and null hypotheses of study, 148–149 Solow, R., 224, 276 SSA, see Sub-Saharan Africa Standard Least Squares, 232 Sub-Saharan Africa (SSA), 127 Superior stable configuration, 89 Support, 15

S

T

Saggi, K., 227 Samoilenko, S., 112, 116, 171, 187 Sapounas, I., 227 SAS Enterprise Miner, 94–95, 117, 118, 232–234, 250 Scale homogeneity, 110–111 Schmutzler, A., 227 Semantic homogeneity, 110 Sen, Amartya, 278 Small and Medium Enterprises, of SSA, 184–185, 201 SME, see Small and Medium Enterprises Smith, Adam, 277 Socio-economic impacts of ICT-enabled public value, in SSA, 203–218 data acquisition, 209–210, 210 data analysis, results of, 210–217 methodology to investigate, 208–210 association rule mining, 209, 224, 224–225 cluster analysis, 209, 210, 210–211, 211 data envelopment analysis, 209, 212, 213 decision tree induction, 209, 211, 211 ordinary least squares regression, 209, 212, 213 overview, 203–204 research framework to study, 204–207 ICT capabilities impact, 205–206 NRI framework, 204–205, 205 research questions on study of, 207–208 results of study on, 216–217

Taylor, Frederick, 162 TE, see Transition economies Technology or total factor productivity (TFP), 224, 225 Theory of complementarity, 225–227 Tobit regression, 114 Top-level splits in decision trees induction, 11, 33 Torrisi S., 227 Total telecom services revenues, 229 Transfer function, 114 Transformative capacity of decision-making units, 47, 111 NN model of, 49–50 Transition economies (TE), 221–223, 246 contributing factors, list of averages in absolute values, and levels of split, 236 and corresponding distribution per group, 236 countries categorize under, 227 definition of, 221 depreciation of, 225 domestic markets of, 222 ICT investments to macroeconomic growth of, 225 information technology investment utilization in, 221–240 investment in telecoms in, 223 relative efficiency of utilization of investments by, 232, 239

292

Index

U

W

United Nations Economic and Social Council, 120 Universe of Discourse of Transformation (UoDoT), 169–170 UoDoT, see Universe of Discourse of Transformation

Ward’s Minimum Variance, 233 World Bank, 221, 227, 246 World Development Indicators database, 227, 234 Worldwide Governance Indicators (WGI), 207 Worthington, A. C., 171 Wulong, Gu., 227

V

Y

Variable returns to scale (VRS), 98, 99, 117, 231 Farrel Input-Saving Measure of Efficiency, 243 relative efficiencies of DMU, 252 VRS, see Variable returns to scale

Yearbook of Statistics (2004), 227, 234, 247

Z Zheng, Z., 157 Zinovyeva, N., 227